Transformers

Datasets Features

Agent traces

• Parse Agent traces messages for SFT using teich by @lhoestq in https://github.com/huggingface/datasets/pull/8232

  • Agent traces from claude_code/pi/codex and others can now be loaded with load_dataset
  • Using the teich library (new optional dependency), traces are parsed to messages to enable training on traces using e.g. trl
  • Load the data:

  >>> from datasets import load_dataset
  >>> ds = load_dataset("lhoestq/agent-traces-example", split="train")
  >>> ds[0]["messages"]
  [{'role': 'user', 'content': 'Download a random dataset from Hugging Face, use DuckDB to inspect it, and come back with a short report about it. Be concise and include: dataset name, what files/format you found, row count or rough size if you can determine it,...'
   ...]

  • Train on agent traces:

  trl sft --dataset-name lhoestq/agent-traces-example ...

  • find all the Agent traces datasets on HF here: https://huggingface.co/datasets?format=format:agent-traces&sort=trending

Next-level shuffling in streaming mode

• Use multiple input shards for shuffle buffer by @lhoestq in https://github.com/huggingface/datasets/pull/8194

  ds = load_dataset(..., streaming=True)
  ds = ds.shuffle(seed=42)
  # or configure local buffer shuffling manually, default is:
  ds = ds.shuffle(seed=42, buffer_size=1000, max_buffer_input_shards=10)

  before👎: <img width="1061" height="529" alt="image" src="https://github.com/user-attachments/assets/dd85a11b-72fc-474c-8faa-33eda779a4fb" />

  after✨: <img width="1057" height="523" alt="image" src="https://github.com/user-attachments/assets/77acc8d3-61d4-4e57-bc87-43fcd3c2e8f3" />

  toy example comparison

  from datasets import IterableDataset
  
  ds = IterableDataset.from_dict({"i": range(123_456_789)}, num_shards=1024)
  ds = ds.shuffle(seed=42)
  
  print("Cold start ids:")
  print(list(ds.take(10)["i"]))
  print("Nominal regime ids:")
  print(list(ds.skip(10_000).take(10)["i"]))

  before👎:

  Cold start ids:
  [6148853, 6149537, 6149418, 6149202, 6149197, 6149622, 6148849, 6149461, 6148965, 6148858]
  Nominal regime ids:
  [6149537, 6149418, 6149202, 6149197, 6149622, 6148849, 6149461, 6148965, 6148858, 6149290]

  after✨:

  Cold start ids:
  [7836668, 9283505, 95847927, 482299, 9283471, 482341, 112003312, 59920157, 43764666, 95847871]
  Nominal regime ids:
  [9283505, 95847927, 482299, 9283471, 482341, 112003312, 59920157, 43764666, 95847871, 16758448]

  Note: ds.state_dict() and ds.load_state_dict() are still supported for this improved shuffling :) enabling dataset checkpointing

  Note 2: it uses threads to fetch the first examples in parallel from the input shards

  Note 3: This is a BREAKING CHANGE: the default shuffling mechanism now uses multiple input shards. You can get the old mechanism by passing max_buffer_input_shards=1 to IterableDataset.shuffle()

New batching features for robotics datasets

• Add batch(by_column=...) by @lhoestq in https://github.com/huggingface/datasets/pull/8172

  from datasets import Dataset
  
  ds = Dataset.from_dict({"episode": [0] * 10 + [1] * 10, "frame": list(range(10)) * 2})
  # ds = ds.to_iterable_dataset()
  ds = ds.batch(by_column="episode")
  for x in ds:
      print(x)
  # {'episode': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 'frame': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]}
  # {'episode': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 'frame': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]}

New supported formats

• Add Apache Iceberg format support by @frankliee in https://github.com/huggingface/datasets/pull/8148
• feat: add TsFile (Apache IoTDB) packaged builder with per-device wide format by @JackieTien97 in https://github.com/huggingface/datasets/pull/8160
• feat: add 3D mesh support and MeshFolder builder by @Vinay-Umrethe in https://github.com/huggingface/datasets/pull/8055
• Add .conll / .conllu dataset format loader (CoNLL-2003 / 2000 / U) by @CrypticCortex in https://github.com/huggingface/datasets/pull/8219

Other improvements and bug fixes

• Pass library_name/version to HfApi in dataset push and delete paths by @davanstrien in https://github.com/huggingface/datasets/pull/8161
• Fix storage_options lookup for streaming Lance datasets by @ericjaebeom in https://github.com/huggingface/datasets/pull/8166
• add agent trace prompt, sent_at, count fields by @cfahlgren1 in https://github.com/huggingface/datasets/pull/8163
• fix: add num_proc argument to Dataset.to_sql by @EricSaikali in https://github.com/huggingface/datasets/pull/7791
• Support fsspec 2026.4.0 by @lhoestq in https://github.com/huggingface/datasets/pull/8175
• Fix Parquet streaming hangs at the end of script by @lhoestq in https://github.com/huggingface/datasets/pull/8176
• ClassLabel docs: Correct value for unknown labels by @l-uuz in https://github.com/huggingface/datasets/pull/7645
• fix parquet reshard by @lhoestq in https://github.com/huggingface/datasets/pull/8193
• Fix parquet columns arg by @lhoestq in https://github.com/huggingface/datasets/pull/8210
• update readme by @lhoestq in https://github.com/huggingface/datasets/pull/8208
• update single seg repos in ci by @lhoestq in https://github.com/huggingface/datasets/pull/8213
• Fix single lance file form pylance 7.0 by @lhoestq in https://github.com/huggingface/datasets/pull/8225
• fix(map): fix progress bar exceeding total when load_from_cache_file=False by @Nitin-Rajasekar in https://github.com/huggingface/datasets/pull/8170
• fix: embed_external_files=True for mesh support by @Vinay-Umrethe in https://github.com/huggingface/datasets/pull/8224
• Fix iterable skip over full Arrow blocks by @my17th2 in https://github.com/huggingface/datasets/pull/8236
• Keep None as a real null in Json() columns instead of the string "null" by @adityasingh2400 in https://github.com/huggingface/datasets/pull/8231
• Support composed splits in streaming datasets by @lanarkite99 in https://github.com/huggingface/datasets/pull/8220

New Contributors

• @ericjaebeom made their first contribution in https://github.com/huggingface/datasets/pull/8166
• @EricSaikali made their first contribution in https://github.com/huggingface/datasets/pull/7791
• @l-uuz made their first contribution in https://github.com/huggingface/datasets/pull/7645
• @CrypticCortex made their first contribution in https://github.com/huggingface/datasets/pull/8219
• @frankliee made their first contribution in https://github.com/huggingface/datasets/pull/8148
• @Vinay-Umrethe made their first contribution in https://github.com/huggingface/datasets/pull/8055
• @Nitin-Rajasekar made their first contribution in https://github.com/huggingface/datasets/pull/8170
• @JackieTien97 made their first contribution in https://github.com/huggingface/datasets/pull/8160
• @my17th2 made their first contribution in https://github.com/huggingface/datasets/pull/8236
• @adityasingh2400 made their first contribution in https://github.com/huggingface/datasets/pull/8231
• @lanarkite99 made their first contribution in https://github.com/huggingface/datasets/pull/8220

Full Changelog: https://github.com/huggingface/datasets/compare/4.8.5...5.0.0

There was a big bug in the model conversion of models related to clip, this affected models like sam3 and others. Please make sure to update 🙏

• Fix conversion for clip models by @zucchini-nlp (#46406)

Full Changelog: https://github.com/huggingface/transformers/compare/v5.10.1...v5.10.2

v5.10.0 was yanked as we publish on a corrupted branch. Sorry everyone, this happens when we rush a release!!!

New Model additions

Gemma4 unified+ Gemma4 MTP

<img width="2000" height="400" alt="image" src="https://github.com/user-attachments/assets/5e3ee940-f78d-4343-ac7a-889930800aa6" />

Gemma 4 12B Unified is an encoder-free multimodal model with pretrained and instruction-tuned variants. Unlike standard Gemma 4, which uses dedicated encoder towers, Gemma 4 12B Unified projects raw inputs directly into the language model's embedding space through lightweight linear pipelines. This results in a simpler architecture while maintaining strong multimodal performance.

Key differences from standard Gemma 4:

• No Vision Tower: Raw pixel patches are projected directly into LM space via a Dense + LayerNorm pipeline with factorized 2D positional embeddings, replacing the vision encoder.
• No Audio Tower: Raw 16 kHz waveform samples are chunked into fixed-length frames and projected through a simple RMSNorm → Linear pipeline, replacing the mel spectrogram + Conformer encoder.
• Shared Multimodal Pipeline: Both vision and audio use the same Gemma4UnifiedMultimodalEmbedder (RMSNorm → Linear) for the final projection to text hidden space.

You can find the original Gemma 4 12B Unified checkpoints under the Gemma 4 release.

• who needs encoders? (#46385) by @douglas-reid @sgerrard @vasqu @molbap

Sapiens2

Sapiens2 is a family of high-resolution vision transformers pretrained on ~1 billion curated human images, designed for human-centric computer vision tasks including pose estimation, body-part segmentation, surface normal estimation, and pointmap estimation. The models scale from 0.4B to 5B parameters and train at native 1K resolution, with hierarchical 4K variants for extended spatial reasoning. Sapiens2 achieves substantial improvements over its predecessor with +4 mAP in pose estimation, +24.3 mIoU in body-part segmentation, and 45.6% error reduction in normal estimation.

Links: Documentation | Paper

• Add Sapiens2 Model (#45919) by @guarin in #45919

DeepSeek-OCR-2

DeepSeek-OCR-2 is an OCR-specialized vision-language model built on a distinctive architecture that combines a SAM ViT-B vision encoder with a Qwen2 hybrid attention encoder, connected through an MLP projector to a DeepSeek-V2 Mixture-of-Experts (MoE) language model. The model features a hybrid attention mechanism that applies bidirectional attention over image tokens and causal attention over query tokens, enabling efficient and accurate document understanding. It supports both plain OCR tasks and grounding capabilities with coordinate-aware output for document conversion to markdown format.

Links: Documentation

• Add Deepseek-OCR-2 model (#45075) by @thisisiron in #45075

Mellum

Mellum is a code-focused Mixture-of-Experts language model developed by JetBrains. It is derived from the Qwen3-MoE architecture with per-layer-type RoPE and interleaved sliding window attention. The model has 12B total parameters with 2.5B active parameters per token, using 64 routed experts with 8 activated per token across 28 layers.

Links: Documentation

• feat: Add support for JetBrains' Mellum v2 code generation model (#46112) by @shadeMe in #46112

Breaking changes

The Gemma4 vision pooler now casts inputs to float32 before scaling to prevent float16 overflow (inf saturation) with large checkpoints, which may cause minor numerical differences in outputs for users running Gemma-4 vision models in float16.

• 🚨 Fix float16 overflow in Gemma4 vision pooler (#46277) by @Bluear7878

Audio Language Models (ALMs) now have a dedicated base model class without a language modeling head, aligning them with the design of Vision Language Models (VLMs); users relying on the previous model class structure should update their code to use the new base model class where appropriate.

• 🚨 [ALM] Add base model without head (#45534) by @eustlb

Parallelization

This release includes numerous bug fixes for model parallelism across multiple models (Gemma4, AltCLIP, ChineseClip, Blip-2, Whisper, Ovis2, Moshi) and parallel execution strategies, including fixes for tensor parallelism (TP), expert parallelism (EP), beam search under model parallel settings, and loss over-counting under TP/EP configurations. The continuous batching manager was also reworked for clearer control flow and improved TP race condition handling, and FSDP initialization via from_pretrained was introduced.

• Fix dsv4 dequant + tp/ep (#46378) by @IlyasMoutawwakil in [#46378]
• [CB] [Major] Rework manager to have clearer control flow + handle TP (#46070) by @remi-or in [#46070]
• fix series of bugs for model parallel beam search (#46280) by @kaixuanliu in [#46280]
• Fix model parallel issue for altclip model and ChineseClip model (#45487) by @kaixuanliu in [#45487]
• Model parallel fix (#46230) by @kaixuanliu in [#46230]
• [Revert] FSDP+Dtensor refactor related changes (#46246) by @vasqu in [#46246]
• Fix model parallel bugs for Gemma4 (#45817) by @kaixuanliu in [#45817]
• init FSDP through from_pretrained (#46102) by @3outeille in [#46102]
• fix model parallel device mismatch issue in create_bidirectional_mask (#46221) by @kaixuanliu in [#46221]
• Trainer.compute_loss: fix loss over-counting under TP and EP-as-TP (#45994) by @AmineDiro in [#45994]
• Fix caching allocator warmup byte estimation for EP model loading (#46149) by @sywangyi in [#46149]

Cache

Fixed a regression in encoder-decoder cache initialization where the decoder config was incorrectly applied to the cross-attention cache, and resolved a RuntimeError caused by buffer size limits when warming up the cache on MPS devices. Additional test infrastructure improvements were made to support read-only cache environments used in CI.

• fix: cache warmup RuntimeError on mps (#46239) by @McPatate in [#46239]
• Make more tests work with read-only cache (#46299) by @ydshieh in [#46299]
• Update a test to avoid writing to the default xet cache (#46250) by @ydshieh in [#46250]
• Fix a regression in encoder-decoder generation cache initialization (#46111) by @kaixuanliu in [#46111]

Quantization

Added support for DeepGEMM BF16, mixed FP8/FP4, and MegaMoE quantization via a grouped linear refactor, while fixing two bugs: an FP8 MoE reverse substring issue affecting DSv4 initialization, and a BitsAndBytes 4-bit/8-bit quantization bug that silently dropped chunked tensors from one-to-many weight converters.

• DeepGEMM BF16 + mixed FP8/FP4 + MegaMoE + refactor (#45634) by @IlyasMoutawwakil in [#45634]
• Fix fp8 moe reverse substring (#46265) by @ArthurZucker in [#46265]
• Fix bnb 4bit/8bit quantization drop chunked tensors bug (#46210) by @kaixuanliu in [#46210]

Bugfixes and improvements

• Fix wrong changes produced by style/repo. check bot (#46371) by @ydshieh in [#46371]
• Fix path traversal when saving Bark voice preset embeddings (#46237) by @LinZiyuu in [#46237]
• Pass library_name/version to Hub calls via a shared HfApi (#46318) by @Wauplin in [#46318]
• docs: update ACL Anthology URL in CITATION.cff (#46352) by @irfaan101 in [#46352]
• [docs] contributing (#45465) by @stevhliu in [#45465]
• [docs] Romanian translation of contributing.md, modular_transformers.md, multimodal_processing.md, add_vision_processing_components.md, add_audio_processing_components.md, modeling_rules.md, model_output_tracing.md, auto_docstring.md, testing.md, pr_checks.md and add_new_model.md . (#46345) by @filipinescu in [#46345]
• [docs] xpu continuous batching (#46334) by @stevhliu in [#46334]
• Fix incorrect attribute mapping relationships in GLM MoE DSA Config (#46338) by @Dovis01 in [#46338]
• Fix grammar typos in Whisper documentation (#46336) by @calliec-1223 in [#46336]
• [docs] update num_items_in_batch for causal LMs (#46335) by @stevhliu in [#46335]
• Update compressed tensors minimum version (#46342) by @SunMarc in [#46342]
• Fix _is_package_available reporting available without a version (#46125) by @blipbyte in [#46125]
• remove sec (#46346) by @ydshieh in [#46346]
• fix: include transitive relative imports when loading from local directory (#46022) by @trducng in [#46022]
• perf(feature_extraction_sequence): skip re-splitting already-batched numpy arrays in pad() (#46329) by @Anai-Guo in [#46329]
• [Zamba] Support attn_implementation dispatch (#46317) by @YangKai0616 in [#46317]
• Fix TestAppRoutes test failures caused by deprecated asyncio.get_event_loop() on Python 3.10+ (#46340) by @ydshieh in [#46340]
• [Qwen3VL] Fix video token placeholder: use self.video_token instead of hardcoded "<|placeholder|>" (#46296) by @kpal002 in [#46296]
• chore(linter): fixes for rule 16 (#46023) by @tarekziade in [#46023]
• [docs] Romanian translation of weightconverter.md, models.md, custom_models.md, monkey_patching.md, fusion_mapping.md, how_to_hack_models.md, model_sharing.md and serialization.md. (#46309) by @filipinescu in [#46309]
• Normalize CUDA OOM errors when comparing commit failures in check_bad_commit (#46322) by @ydshieh in [#46322]
• Fix unhandled exception noise from background safetensors conversion thread (#45752) by @dhruv7477 in [#45752]
• Add Expectations for pipeline token classification tests (#46151) by @kaixuanliu in [#46151]
• [docs] fix auto-add release dates (#46283) by @zucchini-nlp in [#46283]
• Separate pip command syntax for notebook and CLI tabs in Quickstart (#46243) by @pvelayudhan in [#46243]
• Romanian translation of README.md, index.md, installation.md, _config.py and quicktour.md. (#46166) by @filipinescu in [#46166]
• Fall back to flat kwarg when modality dict is passed without it (#46195) by @Ace3Z in [#46195]
• Fix load_adapter OOM caused by full-model warmup sizing (#46145) by @Yooniel in [#46145]
• Replace assert with raise ImportError for optuna/ray dependency checks (#46263) by @SebTardif in [#46263]
• chore(linter): respect TRF017 modeling rule (#46260) by @tarekziade in [#46260]
• Delete dead code in qwen-vl series (#45827) by @zucchini-nlp in [#45827]
• qa: fix ty caching and align CI with local run (#46278) by @tarekziade in [#46278]
• Guard DeviceMesh import in continuous batching (#46205) by @danyalahmed1995 in [#46205]
• Processor compatibility with vLLM (#46258) by @zucchini-nlp in [#46258]
• Fix PR CI workflow cancellation condition (#46276) by @ydshieh in [#46276]
• [fix] toctree (#46106) by @stevhliu in [#46106]
• add more generic support for distributed trainer tests (#46109) by @kaixuanliu in [#46109]
• add XPU Expectations for florence2 and lfm2_vl model test (#46275) by @kaixuanliu in [#46275]
• Fix StaticCache building an empty layer list when num_kv_shared_layers == 0 (#46235) by @tengomucho in [#46235]
• Fix inverted assertion in remove_handler (#46227) by @SebTardif in [#46227]
• [ShieldGemma2] Support attn_implementation dispatch (#46069) by @YangKai0616 in [#46069]
• [Gemma4] Replace one-hot matmul with F.embedding in position embeddings (#46176) by @Sriniketh24 in [#46176]
• fix: kosmos2.5: properly expand embeddings table (#45835) by @nunq in [#45835]
• find pytest launch error in torch 2.13.0.dev20260526 (#46252) by @sywangyi in [#46252]
• [Test][Kosmos2.5] Add XPU expectations for integration tests (#46135) by @YangKai0616 in [#46135]
• Support FA2 flash_attn_with_kvcache for XPU continuous batching (#46028) by @YangKai0616 in [#46028]
• [Configs] Fix layer type validation to include its mlp counterpart (#46220) by @vasqu in [#46220]
• Fix num_items_in_batch over-counting for causal LM losses (#46204) by @qgallouedec in [#46204]
• RF-DETR doc fixes (#46244) by @merveenoyan in [#46244]
• Use main instead of commit SHA for now (#46241) by @ydshieh in [#46241]
• Enable push event (to main) for PR CI workflow (#46240) by @ydshieh in [#46240]
• fix(hrm_text): Add XPU Expectations for tests (#46214) by @kaixuanliu in [#46214]
• [deepseek_v4] keep hc_head / sinks / position_bias in fp32 (#46198) by @ArthurZucker in [#46198]
• Fix FSDP2 and distributed checkpointing imports for older PyTorch versions (#46141) by @ryota-komatsu in [#46141]
• Fix Gemma4 Array Mask Indexing (#46203) by @petecao in [#46203]
• utils: handle flash_attn missing from importlib packages_distributions without crashing (#45524) by @SAY-5 in [#45524]
• [AMD CI] revert AMD mi325 hf-workflows ref from SHA back to @main (#46213) by @Abdennacer-Badaoui in [#46213]
• [GLM-4.6V] Update with GLM-GA Processor (#46184) by @zRzRzRzRzRzRzR in [#46184]
• update xpu expectation for falcon mamba (#46086) by @sywangyi in [#46086]
• chore: enable Dependabot weekly GitHub Actions bumps (#46157) by @hf-dependantbot-rollout[bot] in [#46157]
• Fix Gemma4 use_bidirectional_attention="all" mask behavior (#46079) by @oliverholworthy in [#46079]
• Fix loading with only 1 device or distributed config (#46197) by @Cyrilvallez in [#46197]
• Fix TypeError on list-typed ignore_keys_at_rope_validation in RoPE config (#46142) by @Charly21r in [#46142]
• Support XPU autocast dtype fallback for FlashAttention (#46199) by @YangKai0616 in [#46199]
• Fix path traversal when saving named chat templates (#46191) by @LinZiyuu in [#46191]
• Fix is_last off-by-one in MaskGenerationPipeline for partial batches (#46136) by @J3r3myPerera in [#46136]
• Fix wrong variable in check_model_type isinstance check (#46080) by @SebTardif in [#46080]
• Enable passing kwargs through RoFormer models (#46171) by @ir2718 in [#46171]
• Update cohere2_moe tp_plan (#46189) by @Cyrilvallez in [#46189]
• Update release tool (#46193) by @Cyrilvallez in [#46193]
• [loading] Fix base_model_prefix issues in conversions (#46067) by @Cyrilvallez in [#46067]
• Bump dev version (#46188) by @Cyrilvallez in [#46188]
• Update self-comment-ci (#46137) by @guarin in [#46137]

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @filipinescu
  • [docs] Romanian translation of contributing.md, modular_transformers.md, multimodal_processing.md, add_vision_processing_components.md, add_audio_processing_components.md, modeling_rules.md, model_output_tracing.md, auto_docstring.md, testing.md, pr_checks.md and add_new_model.md . (#46345)
  • [docs] Romanian translation of weightconverter.md, models.md, custom_models.md, monkey_patching.md, fusion_mapping.md, how_to_hack_models.md, model_sharing.md and serialization.md. (#46309)
  • Romanian translation of README.md, index.md, installation.md, _config.py and quicktour.md. (#46166)
• @remi-or
  • [CB] [Major] Rework manager to have clearer control flow + handle TP (#46070)
• @thisisiron
  • Add Deepseek-OCR-2 model (#45075)
• @kaixuanliu
  • Add Expectations for pipeline token classification tests (#46151)
  • fix series of bugs for model parallel beam search (#46280)
  • add more generic support for distributed trainer tests (#46109)
  • add XPU Expectations for florence2 and lfm2_vl model test (#46275)
  • Fix model parallel issue for altclip model and ChineseClip model (#45487)
  • Model parallel fix (#46230)
  • fix(hrm_text): Add XPU Expectations for tests (#46214)
  • Fix model parallel bugs for Gemma4 (#45817)
  • Fix bnb 4bit/8bit quantization drop chunked tensors bug (#46210)
  • fix model parallel device mismatch issue in create_bidirectional_mask (#46221)
  • Fix a regression in encoder-decoder generation cache initialization (#46111)
• @shadeMe
  • feat: Add support for JetBrains' Mellum v2 code generation model (#46112)
• @vasqu
  • [Revert] FSDP+Dtensor refactor related changes (#46246)
  • [Configs] Fix layer type validation to include its mlp counterpart (#46220)
• @zRzRzRzRzRzRzR
  • [GLM-4.6V] Update with GLM-GA Processor (#46184)
• @eustlb
  • 🚨 [ALM] Add base model without head (#45534)

New Model additions

Cohere2Moe

Command A+ is a Mixture-of-Experts (MoE) language model from Cohere that features a hybrid attention pattern combining sliding window and full attention layers. The model incorporates both shared and routed experts and supports a very large context window for processing extensive text sequences.

Links: Documentation

• Add new cohere2_moe model (#46115) by @Cyrilvallez in #46115

Parakeet tdt (#44171)

• Parakeet tdt (#44171) by @lmaksym

HRM-Text

HRM-Text is an improved autoregressive language-modeling variant of the Hierarchical Reasoning Model (HRM) that uses a hierarchical recurrent forward pass with two transformer stacks - one for slow, abstract planning (H) and one for fast, detailed computation (L) - reused inside a nested recurrence. It features PrefixLM attention where instruction tokens attend bidirectionally while response tokens attend causally, per-head sigmoid output gates, and parameterless RMSNorm. The model is designed as a base language model without instruction tuning or chat templates.

Links: Documentation | Paper

• Add hrm text (#46025) by @abcd1927 in #46025

Breaking changes

The text_embeds input for SAM3, EdgeTAM, and SAM3-Lite-Text models now expects full text embeddings instead of just pooler outputs, aligning with other models in the library — users must update their inputs accordingly.

• 🚨Fix memory leaks caused by lru decorators in vision models (#45922) by @yonigozlan

Audio

Audio support was expanded with the addition of AudioFlamingoNext model checkpoints and improved compilability of audio/vision encoders via standalone pure functions. Additional improvements include better error messaging when loading audio from video files and new documentation for audio/video processors.

• user friendly error when loading audio from video (#45221) by @eustlb in [#45221]
• [docs] adding audio/video processors (#45795) by @stevhliu in [#45795]
• Support Audio Flamingo Next checkpoints (#44830) by @lashahub in [#44830]
• Extract dynamic vision/audio tensors into standalone pure functions (#45396) by @IlyasMoutawwakil in [#45396]

Generation

Fixed generation issues including inputs_embeds and per_layer_inputs handling for Gemma4, an AttributeError in RAG's generate() caused by missing config fields, and flaky VLM generation tests by blocking special image tokens during sampling.

• Fix Gemma4 generation from inputs_embeds and per_layer_inputs (#46049) by @Cyrilvallez in [#46049]
• Fix AttributeError in RAG generate() for missing config fields (#46035) by @Sriniketh24 in [#46035]
• Block image_start/end_token_id in generation test sampling (#45914) by @Rocketknight1 in [#45914]

Bugfixes and improvements

• Remove mask visualization tool from masking_utils.py (#46066) by @Cyrilvallez in [#46066]
• fix: owned_by field in GET /v1/models returns list instead of string (#46006) by @nileshpatil6 in [#46006]
• [CB] Remove OpenTelemetry (#45984) by @remi-or in [#45984]
• docs(readme): use canonical huggingface.co domain in prose links (#46042) by @kiwigitops in [#46042]
• Fix remaining RAG doc examples that crash on current transformers (#46044) by @Sriniketh24 in [#46044]
• Init the actual tensor, not a copy (#46030) by @Rocketknight1 in [#46030]
• docs: sync legacy ACL anthology URLs and update metrics across i18n READMEs (#46027) by @irfaan101 in [#46027]
• [MultimodalLM] add language_model to the get/set_input_embeddings logic (#46029) by @eustlb in [#46029]
• [HRM Text] Add integration tests (#46033) by @vasqu in [#46033]
• hy_v3: add XPU expectations (#45858) by @kaixuanliu in [#45858]
• exaone4_5: add XPU expectations (#45890) by @kaixuanliu in [#45890]
• hyperclovax: add XPU Expectations for CI test (#45926) by @kaixuanliu in [#45926]
• chore(ci): remove dead env vars from circleci-failure-summary-comment.yml (#45972) by @XciD in [#45972]
• [CB] [Major] Add tensor paralellism (#45821) by @remi-or in [#45821]
• docs: update models architecture count and sync ACL anthology URLs (#46001) by @irfaan101 in [#46001]
• bugfix(ci): avoid E2BIG in pr_slow_ci_suggestion (#45983) by @tarekziade in [#45983]
• RFDetr - use correct Roboflow org for release (#45946) by @sbucaille in [#45946]
• docs: Fix formatting issues in weightconverter.md (#45988) by @ArjunSrivastava1 in [#45988]
• Fix colqwen2 test (#45981) by @IlyasMoutawwakil in [#45981]
• Fix M-RoPE device mismatch in Qwen3VL family under FSDP2 CPU offload (#45861) by @jamesbraza in [#45861]
• [docs] chat template prefill (#45947) by @stevhliu in [#45947]
• [docs] decode fast path (#45899) by @stevhliu in [#45899]
• fix: restore _attn_implementation and fix request offset in generate_batch() (#45943) by @sergiopaniego in [#45943]
• Expose per_layer_inputs for every Gemma4 variants (#45927) by @Cyrilvallez in [#45927]
• chore: update benchmark_v2.yml (#45966) by @hf-security-analysis[bot] in [#45966]
• fix(ci): set persist-credentials: false on actions/checkout and close remaining template injection findings (#45964) by @XciD in [#45964]
• chore(ci): set default workflow permissions to contents: read (#45961) by @XciD in [#45961]
• fix(ci): remove template injection on pull_request_target workflows (#45956) by @XciD in [#45956]
• chore(ci): pin all GitHub Actions and reusable workflows by SHA (#45955) by @XciD in [#45955]
• [docs] ALMModelTest (#45900) by @stevhliu in [#45900]
• Enhance apply_chat_template to support custom field prefilling (reasoning_content, thinking, etc.) (#45896) by @Mamiglia in [#45896]
• BUGFIX: Support hubert models that don't have conv_pos_batch_norm configured (#45921) by @igordertigor in [#45921]
• Revert 45777 (#45942) by @Rocketknight1 in [#45942]
• pass the otel secrets (#45933) by @tarekziade in [#45933]
• Add initial torch_tpu backend support (#45918) by @tengomucho in [#45918]
• [CB] Hide activation footprint by using the CUDA graph pool (#45911) by @remi-or in [#45911]
• Require input_ids for repetition penalty (#45389) by @ruben-aghayan in [#45389]
• Fix undefined 'input' variable (#45895) by @fullyz in [#45895]
• Fix post processing RF-DETR (#46041) by @yonigozlan (direct commit on v5.9.0)
• [loading] Free up tensors faster inside ConversionOps (#46110) by @Cyrilvallez (direct commit on v5.9.0)
• Add new cohere2_moe model (#46115) by @Cyrilvallez (direct commit on v5.9.0)
• Fix cohere2 tp_plan for release by @Cyrilvallez (direct commit on v5.9.0)
• Release v5.9.0 by @Cyrilvallez (direct commit on v5.9.0)

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @lmaksym
  • Parakeet tdt (#44171)
• @eustlb
  • user friendly error when loading audio from video (#45221)
  • [MultimodalLM] add language_model to the get/set_input_embeddings logic (#46029)
• @remi-or
  • [CB] Remove OpenTelemetry (#45984)
  • [CB] [Major] Add tensor paralellism (#45821)
  • [CB] Hide activation footprint by using the CUDA graph pool (#45911)
• @abcd1927
  • Add hrm text (#46025)

This release is mainly to fix the Deepseek V4 integration!!!

<img width="714" height="774" alt="image" src="https://github.com/user-attachments/assets/0d85e891-a0ff-436e-a9d4-b6633096f2b5" />

• [fix] Add fatal_error to ContinuousBatchingManager so the serving... by @qgallouedec, @remi-or
• Fix WeightConverter regex incorrectly matching shared_experts as experts by @silencelamb, @claude
• Fix deepseek v4 by @ArthurZucker (#45892)
• Deepseek v4 csa mask collapse by @ArthurZucker, @Sawyer117 (#45928)

Release v5.8.0

New Model additions

DeepSeek-V4

<img width="6604" height="3574" alt="image" src="https://github.com/user-attachments/assets/4c0fdb29-f770-463c-a97b-d24438896a4c" />

DeepSeek-V4 is the next-generation MoE (Mixture of Experts) language model from DeepSeek that introduces several architectural innovations over DeepSeek-V3. The architecture replaces Multi-head Latent Attention (MLA) with a hybrid local + long-range attention design, swaps residual connections for Manifold-Constrained Hyper-Connections (mHC), and bootstraps the first few MoE layers with a static token-id → expert-id hash table. This implementation covers DeepSeek-V4-Flash, DeepSeek-V4-Pro, and their -Base pretrained variants, which share the same architecture but differ in width, depth, expert count and weights.

Links: Documentation | Paper

• Add DeepSeek V4 (#45643) by @ArthurZucker in #45643

Gemma 4 Assistant

<img width="2000" height="400" alt="image" src="https://github.com/user-attachments/assets/02c79b0b-a172-4495-b09d-a6a4b625ee66" />

Gemma 4 Assistant is a small, text-only model that enables speculative decoding for Gemma 4 models using the Multi-Token Prediction (MTP) method and associated candidate generator. The model shares the same Gemma4TextModel backbone as other Gemma 4 models but uses KV sharing throughout the entire model, allowing it to reuse the KV cache populated by the target model and skip the pre-fill phase entirely. This architecture includes cross-attention to make the most of the target model's context, allowing the assistant to accurately predict more drafted tokens per drafting round.

Links: Documentation

• First model (#45788) by @SindhuRaghuram97 in #45788

GraniteSpeechPlus

<img width="1310" height="930" alt="image" src="https://github.com/user-attachments/assets/94fc3730-742c-4b9e-ab6a-ed2e5c75d0bf" />

Granite Speech Plus is a variant of Granite Speech that enhances the projector by consuming the concatenation of the encoder's final hidden states with an arbitrary subset of its intermediate hidden states along the feature dimension. It is a multimodal speech-to-text model that can transcribe audio, provide speaker annotation and word level timestamps by responding to text prompts. The model inherits the same architecture components as Granite Speech including the speech encoder, query transformer projector, language model, and optional LoRA adapter.

Links: Documentation

• Support for a new Granite-Speech-Plus model (#45695) by @zvik in #45695

Granite4Vision

Granite Vision 4.1 is a vision-language model from IBM Research designed for enterprise-grade document data extraction. It specializes in chart extraction (Chart2CSV, Chart2Summary, Chart2Code), table extraction (JSON, HTML, OTSL), and semantic key-value pair extraction. The model builds on LLaVA-NeXT with architectural innovations including SigLIP2 Vision Encoder, Window Q-Former Projectors, and DeepStack Feature Injection with 8 vision-to-LLM injection points.

Links: Documentation

• Add Granite 4.1 Vision (granite4_vision) (#45597) by @artem-spector in #45597

EXAONE-4.5

<img width="3840" height="2160" alt="image" src="https://github.com/user-attachments/assets/55eb732d-f9da-4f97-8226-2cd3f6476ca0" />

EXAONE 4.5 is the first open-weight vision language model developed by LG AI Research, integrating a dedicated visual encoder into the existing EXAONE 4.0 framework to expand multimodal capabilities. The model features 33 billion parameters in total, including 1.2 billion parameters from the vision encoder, and achieves competitive performance in general benchmarks while outperforming similar-sized models in document understanding and Korean contextual reasoning. It builds on EXAONE 4.0 with key enhancements including an expanded vocabulary of 153,600 tokens, support for up to 256K token context windows, and a Multi-Token Prediction (MTP) mechanism.

Links: Documentation | Paper | Blog Post

• Add EXAONE 4.5 implementations (#45471) by @nuxlear in #45471

PP-FormulaNet

PP-FormulaNet-L and PP-FormulaNet_plus-L are lightweight models designed for table structure recognition, focusing on accurately recognizing table structures in documents and natural scenes. The models are part of the SLANet series and can be used for image-to-text tasks, specifically for detecting and processing mathematical formulas and table structures from images.

Links: Documentation

• [Model] Add PP-FormulaNet Model Support (#45626) by @zhang-prog in #45626

Breaking changes

Apex integration has been removed from the library (including RMSNorm usage in T5 and related models), so users relying on Apex for mixed precision or fused ops should migrate to PyTorch's native equivalents instead.

• 🚨 Get rid of most Apex references (#45723) by @Rocketknight1

Tokenization

Fixed tokenizer mapping issues for DeepSeek R1 distilled (Qwen2) and DeepSeek OCR models, and resolved a significant performance regression in PreTrainedTokenizer.convert_ids_to_tokens where skip_special_tokens=True was rebuilding the special token set on every iteration, resulting in a ~300x speedup for that code path.

• deepseek r1 distilled tokenizer fix for qwen2 mapping (#45741) by @itazap in [#45741]
• DeepSeek OCR specifies an incorrect tokenizer class on the Hub (#45739) by @hmellor in [#45739]
• PythonBackend slow tokenizer convert_ids_to_tokens fix (#45728) by @i3hz in [#45728]

Bugfixes and improvements

• fix: correct spelling in continuous_api docstring (#45749) by @Dhruv908615 in [#45749]
• Fix link to modular transformers documentation (#45746) by @SangbumChoi in [#45746]
• Gemma4: fix failed test cases (#45568) by @kaixuanliu in [#45568]
• Fix CI: Allow more artifacts to be download in CI (#45785) by @ydshieh in [#45785]
• Add concurrency to PR CI workflow file (pr-ci-caller.yml) (#45786) by @ydshieh in [#45786]
• Reorder decorators for autodoc and dataclass (#45702) by @zucchini-nlp in [#45702]
• Unwrap text_config in AutoModelFor*.from_config (#45770) by @jamesbraza in [#45770]
• fix: Added Mps support in float fallback backends list (#45687) by @rigen1048 in [#45687]
• Github Actions PR CI (caller) (#45476) by @ydshieh in [#45476]
• make sure we call check_auto in CI (#45775) by @tarekziade in [#45775]
• Fix auto mapping script (#45774) by @Cyrilvallez in [#45774]
• [MINISTRAL3] Fix conversion script yarn's apply_scale support. (#45744) by @juliendenize in [#45744]
• [nemotron_h] respect _no_reinit flag on dt_bias and out_proj.weight (#45591) by @vai-minzhou in [#45591]
• fix(utils): Resolve backbone utils test regressions (#45594) by @harshaljanjani in [#45594]
• [CB] Better overall script and decode bucketting (#45653) by @remi-or in [#45653]
• [docs] model testing (#45152) by @stevhliu in [#45152]
• update dev (#45726) by @vasqu in [#45726]
• Doc translate to Persian(farsi) (#45664) by @zeoses in [#45664]
• [OAI Privacy Filter] Add integration test (#45725) by @vasqu in [#45725]
• Speedup Qwen2VLImageProcessor (#45719) by @lgeiger in [#45719]
• Remove dead beam-search dummies from dummy_pt_objects.py (#45722) by @jw9603 in [#45722]
• chore(typing): add ty type checking for 10 utility files (#45703) by @moonbogi in [#45703]
• Llama3 video fix (#45040) by @sywangyi in [#45040]
• Fix custom-module copies inheriting read-only permissions (#45686) by @nurpax in [#45686]
• Python code in model docs (#45608) by @zucchini-nlp in [#45608]
• fix failed test cases for blt model (#45596) by @kaixuanliu in [#45596]
• chore(typing): add ty type checking for 3 pipeline files (#45667) by @moonbogi in [#45667]

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @artem-spector
  • Add Granite 4.1 Vision (granite4_vision) (#45597)
• @SindhuRaghuram97
  • First model (#45788)
• @nuxlear
  • Add EXAONE 4.5 implementations (#45471)
• @ArthurZucker
  • Add DeepSeek V4 (#45643)
• @remi-or
  • [CB] Better overall script and decode bucketting (#45653)
• @zhang-prog
  • [Model] Add PP-FormulaNet Model Support (#45626)
• @zvik
  • Support for a new Granite-Speech-Plus model (#45695)

New Model additions

Laguna

<img width="699" height="176" alt="image" src="https://github.com/user-attachments/assets/d3bae269-bea7-4ddf-a53f-d4718befdb17" />

Laguna is Poolside's mixture-of-experts language model family that extends standard SwiGLU MoE transformers with two key innovations. It features per-layer head counts allowing different decoder layers to have different query-head counts while sharing the same KV cache shape, and implements a sigmoid MoE router with auxiliary-loss-free load balancing that uses element-wise sigmoid of gate logits plus learned per-expert bias for router scoring.

Links: Documentation

• Laguna XS.2 implementation (#45673) by @joerowell in #45673

DEIMv2

<img width="2874" height="908" alt="image" src="https://github.com/user-attachments/assets/fc8c59fe-f964-42ce-ae8e-c7fcace9beb7" />

DEIMv2 (DETR with Improved Matching v2) is a real-time object detection model that extends DEIM with DINOv3 features and spans eight model sizes from X to Atto for diverse deployment scenarios. It uses a Spatial Tuning Adapter (STA) for larger variants to convert DINOv3's single-scale output into multi-scale features, while ultra-lightweight models employ pruned HGNetv2 backbones. The unified design achieves superior performance-cost trade-offs, with DEIMv2-X reaching 57.8 AP with only 50.3M parameters and DEIMv2-S being the first sub-10M model to exceed 50 AP on COCO.

Links: Documentation | Paper

• model: Add DEIMv2 to Transformers (#44339) by @harshaljanjani in #44339

Attention

Several attention-related bugs were fixed across multiple models, including a cross-attention cache type error in T5Gemma2 for long inputs, incorrect cached forward behavior in Qwen3.5's gated-delta-net linear attention, and a crash in GraniteMoeHybrid when no Mamba layers are present. Attention function dispatch was also updated to align with the latest model implementations.

• Fix cross-attention cache layer type for T5Gemma2 long inputs (#45540) by @Beichen-Ma in [#45540]
• [Qwen3.5] Fix GDN linear attention multi-token cached forward (#45513) by @kashif in [#45513]
• Fix GraniteMoeHybrid _update_mamba_mask crash on attention-only models (#45514) by @tianhaocui in [#45514]
• Align latest model attention function dispatch (#45598) by @Cyrilvallez in [#45598]

Tokenizers

There was a bug in AutoTokenizer that caused the wrong tokenizer class to be initialized. This caused regressions in models like DeepSeek R1.

• change got reverted (#45680) by @itazap in [#45680]

Generation

Continuous batching generation received several fixes and improvements, including correcting KV deduplication and memory estimation for long sequences (16K+), and removing misleading warnings about num_return_sequences and other unsupported features that were incorrectly firing even when functionality worked correctly. Documentation for per-request sampling parameters was also added.

• generate: drop stale num_return_sequences warning on continuous batching path (#45582) by @joaquinhuigomez in [#45582]
• Remove unnecessary generate warnings (#45619) by @Cyrilvallez in [#45619]
• [CB] Changes for long generation (#45530) by @remi-or in [#45530]
• [docs] per-request sampling params (#45553) by @stevhliu in [#45553]

Kernels

Improved kernel support by fixing configuration reading and error handling for FP8 checkpoints (e.g., Qwen3.5-35B-A3B-FP8), enabling custom expert kernels registered from the HF Hub to be properly loaded, and resolving an incompatibility that prevented Gemma3n and Gemma4 from using the rotary kernel.

• Fix configuration reading and error handling for kernels (#45610) by @hmellor in [#45610]
• Allow for registered experts from kernels hub (#45577) by @winglian in [#45577]
• Gemma3n and Gemma4 cannot use rotary kernel (#45564) by @Cyrilvallez in [#45564]

Bugfixes and improvements

• fixing more typos (#45689) by @vasqu in [#45689]
• [docs] cb memory management (#45587) by @stevhliu in [#45587]
• [docs] cpu offloading (#45660) by @stevhliu in [#45660]
• docs(README_zh-hans): clarify conditions for not using Transformers (#45688) by @GuaiZai233 in [#45688]
• fix padding side issue for fast_vlm tests (#45592) by @kaixuanliu in [#45592]
• Fix x_clip: 8 failed test cases (#45394) by @kaixuanliu in [#45394]
• zero_shot_object_detection ValueError fix for python 3.13 (#45669) by @AnkitAhlawat7742 in [#45669]
• Fix pageable H2D copies in Gated DeltaNet PyTorch fallback (#45665) by @ruixiang63 in [#45665]
• Fix UnboundLocalError in shard_and_distribute_module for replicated parameters (#45675) by @Abdennacer-Badaoui in [#45675]
• [MistralCommonBackend] Soften validation mode and apply_chat_template arguments check (#45628) by @juliendenize in [#45628]
• Fix NameError: PeftConfigLike triggered by PreTrainedModel.__init_subclass__ (#45658) by @qgallouedec in [#45658]
• chore(typing): added modeling_utils to ty (#45425) by @tarekziade in [#45425]
• [gemma4] infer from config instead of hardcoding (#45606) by @eustlb in [#45606]
• Update quants tests (#45480) by @SunMarc in [#45480]
• 🔴🔴🔴 fix: skip clean_up_tokenization for BPE tokenizers in PreTrainedTokenizerFast (#44915) by @maxsloef-goodfire in [#44915]
• Fix colmodernvbert tests (#45652) by @Cyrilvallez in [#45652]
• [CB] [Major] Add CPU request offloading (#45184) by @remi-or in [#45184]
• Fix peft constructors (#45622) by @Cyrilvallez in [#45622]
• chore: speedup modular converter (~30%) (#45046) by @tarekziade in [#45046]
• Fix whisper return language (#42227) by @FredHaa in [#42227]
• Add supports_gradient_checkpointing to NemotronHPreTrainedModel (#45625) by @sergiopaniego in [#45625]
• Raise clear error for problem_type="single_label_classification" with num_labels=1 (#45611) by @gaurav0107 in [#45611]
• CircleCI with torch 2.11 (#45633) by @ydshieh in [#45633]
• chore: bump doc-builder SHA for main doc build workflow (#45631) by @rtrompier in [#45631]
• Allow more artifacts to be download in CI (#45629) by @ydshieh in [#45629]
• chore(qa): split pipeline and add type checking (#45432) by @tarekziade in [#45432]
• Skip failing offloading tests (#45624) by @Cyrilvallez in [#45624]
• fix: compute auxiliary losses when denoising is disabled in D-FINE (#45601) by @Abineshabee in [#45601]
• qa: bumped mlinter and allow local override (#45585) by @tarekziade in [#45585]
• Processing Utils: continue when content is a string (#45605) by @RyanMullins in [#45605]
• SonicMoe (#45433) by @IlyasMoutawwakil in [#45433]
• fix transformers + torchao nvfp4 serialization (#45573) by @vkuzo in [#45573]
• [AMD CI] Fix expectations for Gemma3n (#45602) by @Abdennacer-Badaoui in [#45602]
• [docs] multi-turn tool calling (#45554) by @stevhliu in [#45554]
• Fix AttributeError on s_aux=None in flash_attention_forward (#45589) by @jamesbraza in [#45589]
• do not index past decoded chars with special tokens (#45435) by @itazap in [#45435]
• Update dev version (#45583) by @vasqu in [#45583]
• Update torchao usage for XPU and CPU (#45560) by @jiqing-feng in [#45560]

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @vasqu
  • fixing more typos (#45689)
  • Update dev version (#45583)
• @joerowell
  • Laguna XS.2 implementation (#45673)
• @tarekziade
  • chore(typing): added modeling_utils to ty (#45425)
  • chore: speedup modular converter (~30%) (#45046)
  • chore(qa): split pipeline and add type checking (#45432)
  • qa: bumped mlinter and allow local override (#45585)
• @harshaljanjani
  • model: Add DEIMv2 to Transformers (#44339)
• @remi-or
  • [CB] [Major] Add CPU request offloading (#45184)
  • [CB] Changes for long generation (#45530)

New Model additions

OpenAI Privacy Filter

OpenAI Privacy Filter is a bidirectional token-classification model for personally identifiable information (PII) detection and masking in text. It is intended for high-throughput data sanitization workflows where teams need a model that they can run on-premises that is fast, context-aware, and tunable. The model labels an input sequence in a single forward pass, then decodes coherent spans with a constrained Viterbi procedure, predicting probability distributions over 8 privacy-related output categories for each input token.

Links: Documentation

• [Privacy Filter] Add model (#45580) by @vasqu in #45580

QianfanOCR

Qianfan-OCR is a 4B-parameter end-to-end document intelligence model developed by Baidu that performs direct image-to-text conversion without traditional multi-stage OCR pipelines. It supports a broad range of prompt-driven tasks including structured document parsing, table extraction, chart understanding, document question answering, and key information extraction all within one unified model. The model features a unique "Layout-as-Thought" capability that generates structured layout representations before producing final outputs, making it particularly effective for complex documents with mixed element types.

Links: Documentation | Paper

• add Qianfan-OCR model definition (#45280) by @marvinzh in #45280

SAM3-LiteText

SAM3-LiteText is a lightweight variant of SAM3 that replaces the heavy SAM3 text encoder (353M parameters) with a compact MobileCLIP-based text encoder optimized through knowledge distillation, while keeping the SAM3 ViT-H image encoder intact. This reduces text encoder parameters by up to 88% while maintaining segmentation performance comparable to the original model. The model enables efficient vision-language segmentation by addressing the redundancy found in text prompting for segmentation tasks.

Links: Documentation | Paper

• Add SAM3-LiteText (#44320) by @NielsRogge in #44320

SLANet

SLANet and SLANet_plus are lightweight models designed for table structure recognition, focusing on accurately recognizing table structures in documents and natural scenes. The model improves accuracy and inference speed by adopting a CPU-friendly lightweight backbone network PP-LCNet, a high-low-level feature fusion module CSP-PAN, and a feature decoding module SLA Head that aligns structural and positional information. SLANet was developed by Baidu PaddlePaddle Vision Team as part of their table structure recognition solutions.

Links: Documentation

• [Model] Add SLANet Model Support (#45532) by @zhang-prog in #45532

Breaking changes

The internal rotary_fn is no longer registered as a hidden kernel function, so any code referencing self.rotary_fn(...) within an Attention module will break and must be updated to call the function directly instead.

• 🚨 [Kernels] Fix kernel function registration (#45420) by @vasqu

Serve

The transformers serve command received several enhancements, including a new /v1/completions endpoint for legacy text completion, multimodal support for audio and video inputs, improved tool-calling via parse_response, proper forwarding of tool_calls/tool_call_id fields, a 400 error on model mismatch when the server is pinned to a specific model, and fixes for the response API. Documentation was also updated to cover new serving options such as --compile and --model-timeout.

• Add /v1/completions endpoint (OpenAI legacy completions API) to transformers serve (#44558) by @rain-1 in [#44558]
• Updated the image cache for Paddle models according to the latest API (#45562) by @zhang-prog in [#45562]
• Raise 400 on model mismatch when transformers serve is pinned (#45443) by @qgallouedec in [#45443]
• [serve] Update tool call to switch to parse_response (#45485) by @SunMarc in [#45485]
• Fix response api support (#45463) by @SunMarc in [#45463]
• [serve] Forward tool_calls/tool_call_id in processor inputs (#45418) by @qgallouedec in [#45418]
• refactor(qa): extend extras so ty can run on server modules (#45456) by @tarekziade in [#45456]
• Multimodal serve support (#45220) by @SunMarc in [#45220]
• [docs] transformers serve (#45174) by @stevhliu in [#45174]

Vision

Several vision-related bug fixes were applied in this release, including correcting Qwen2.5-VL temporal RoPE scaling for still images, fixing missing/mismatched image processor backends for Emu3 and BLIP, resolving modular image processor class duplication, and preventing accelerate from incorrectly splitting vision encoders in PeVideo/PeAudioVideo models. Image loading performance was also improved by leveraging torchvision's native decode_image in the torchvision backend, yielding up to ~17% speedup over PIL-based loading.

• Revert "Fix: modular image processors (#45492)" (#45531) by @tarekziade in [#45531]
• Fix: modular image processors (#45492) by @zucchini-nlp in [#45492]
• fix: prevent accelerate from splitting vision encoder by setting no… (#43047) by @<NOT FOUND> in [#43047]
• Fix Qwen2.5-VL temporal RoPE scaling applied to still images (#45330) by @Kash6 in [#45330]
• Use torchvision decode_image to load images in the torchvision backend (#45195) by @yonigozlan in [#45195]
• Fix missing image processors backends (#45165) by @zucchini-nlp in [#45165]

Parallelization

Fixed several bugs affecting distributed training, including silently wrong results or NaN loss with Expert Parallelism, NaN weights on non-rank-0 FSDP processes, and a resize failure in PP-DocLayoutV3; additionally added support for loading adapters with Tensor Parallelism, added MoE to the Gemma4 TP plan, and published documentation for TP training.

• Fix EP: RouterParallel shape, tp_plan property, grouped_mm sentinels (#45473) by @AmineDiro in [#45473]
• Fix NaN weights on non-rank-0 FSDP processes (#45050) by @albertvillanova in [#45050]
• Load adapter with TP (#45155) by @michaelbenayoun in [#45155]
• [docs] tp training (#44613) by @stevhliu in [#44613]
• Fix resize failure caused by zero-sized masks in PP-DocLayoutV3 (#45281) by @zhang-prog in [#45281]
• Add MoE to Gemma4 TP plan (#45219) by @sywangyi in [#45219]

Tokenization

Fixed a docstring typo in streamer classes, resolved a Kimi-K2.5 tokenizer regression and _patch_mistral_regex AttributeError, and patched a streaming generation crash for Qwen3VLProcessor caused by incorrect _tokenizer attribute access. Additional housekeeping included moving the GPT-SW3 instruct tokenizer to an internal testing repo and fixing a global state leak in the tokenizer registry during tests.

• [Doc] Fix 'tokenized' -> 'tokenizer' typo in streamer docstrings (#45508) by @avasis-ai in [#45508]
• Fix Kimi-K2.5 tokenizer regression and _patch_mistral_regex AttributeError (#45359) by @ArthurZucker in [#45359]
• fix(serving): resolve rust tokenizer from ProcessorMixin in streaming generation (#45368) by @sharziki in [#45368]
• [Tokenizers] Move gpt sw3 tokenizer out (#45404) by @vasqu in [#45404]
• fix: leak in tokenizer registry for test_processors (#45318) by @tarekziade in [#45318]

Cache

Cache handling was improved for Gemma4 and Gemma3n models by dissociating KV state sharing from the Cache class, ensuring KV states are always shared regardless of whether a Cache is used. Additionally, the image cache for Paddle models was updated to align with the latest API.

• Align gemma3n cache sharing to gemma4 (#45489) by @Cyrilvallez in [#45489]
• remove cache file from tree (#45392) by @tarekziade in [#45392]
• [gemma4] Dissociate kv states sharing from the Cache (#45312) by @Cyrilvallez in [#45312]

Audio

Audio models gained vLLM compatibility through targeted fixes across several model implementations, while reliability improvements were also made including exponential back-off retries for audio file downloads, a crash fix in the text-to-speech pipeline when generation configs contain None values, and corrected test failures for Kyutai Speech-To-Text.

• feat[vLLM × v5]: Add vLLM compatibility for audio models (#45326) by @harshaljanjani in [#45326]
• http retries on audio file downloads (#45126) by @tarekziade in [#45126]
• fix(testing): Fix Kyutai Speech-To-Text and LongCatFlash test failures on main CI (#44695) by @harshaljanjani in [#44695]
• Fix text-to-speech pipeline crash when generation config contains None values (#45107) by @jiqing-feng in [#45107]

Bugfixes and improvements

• [Privacy Filter] Add model (#45580) by @vasqu in [#45580]
• Add ForSequenceClassification heads for the OLMo family (#45551) by @earino in [#45551]
• Add IndexCache support for GLM5 DSA (#45424) by @louzongzhi in [#45424]
• Fix redundant logic in video processing SmolVLM (#45272) by @yonigozlan in [#45272]
• Fix typos (#45574) by @vasqu in [#45574]
• [Model] Add SLANet Model Support (#45532) by @zhang-prog in [#45532]
• refactor(Dots1): drop Dots1MoE override to pass (inherits from DSV3 MoE) (#45572) by @casinca in [#45572]
• perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models (#45555) by @casinca in [#45555]
• Gemma4 training with text-only samples (#45454) by @zucchini-nlp in [#45454]
• [nemotron_h] Add support for MLP mixers (#44763) by @xenova in [#44763]
• add expert parallelism for gemma-4-26B-A4B-it (#45279) by @sywangyi in [#45279]
• Add full GGUF loading support for GPT‑OSS (fixes #43366, supersedes #43757) latest (#45506) by @sirzechs66 in [#45506]
• Update Gemma4 weight conversion script (#45328) by @RyanMullins in [#45328]
• Move some conversion mappings to PrefixChange (#45567) by @Cyrilvallez in [#45567]
• fix table update versions (#45544) by @tarekziade in [#45544]
• Add disable_mmap kwarg to from_pretrained with hf-mount auto-detection (#45547) by @rtrompier in [#45547]
• fix(DSV3): parity between native DeepseekV3MoE and remote official implementation (#45441) by @casinca in [#45441]
• [modular] Fix modular logic broken in #45045 (#45539) by @Cyrilvallez in [#45539]
• Fix: propagate quantization_config to text sub-config for composite models in AutoModelForCausalLM (#45494) by @lvliang-intel in [#45494]
• T5Gemma2: fix prepare_decoder_input_ids_from_labels (#45516) by @Tokarak in [#45516]
• [Trainer] Add ddp_static_graph option (#45519) by @KeitaW in [#45519]
• Add dtype config options for Four Over Six (#45367) by @jackcook in [#45367]
• [Sam3LiteText] Remove unnecessary modules/configs (#45535) by @yonigozlan in [#45535]
• Fix conditional check for float formatting (#44425) by @qgallouedec in [#44425]
• Fix AMD CI: rebuild torchvision with libjpeg + refresh expectations (#45533) by @Abdennacer-Badaoui in [#45533]
• Reapply modular to examples (#45527) by @Cyrilvallez in [#45527]
• qa: re-run modular converter when the script itself is modified (#45528) by @tarekziade in [#45528]
• [GGUF] Reduce peak RAM usage by casting dequantized tensors early during load (#45386) by @UsamaKenway in [#45386]
• Fix CSM TextToAudioPipeline missing <bos> token (#45525) by @jiqing-feng in [#45525]
• [Conversion Mapping] Small fixups (#45483) by @vasqu in [#45483]
• fix: return empty tuple from import_protobuf_decode_error when protobuf is unavailable (#45486) by @jw9603 in [#45486]
• throw error when conversion required (#45078) by @itazap in [#45078]
• chore: bump doc-builder SHA for PR upload workflow (#45450) by @rtrompier in [#45450]
• xpu output align with cuda in test case (#45526) by @sywangyi in [#45526]
• chore(qa): split out mlinter (#45475) by @tarekziade in [#45475]
• [loading] Clean way to add/remove full parts in checkpoint names (#45448) by @Cyrilvallez in [#45448]
• Fix Zamba2MambaMixer ignoring use_mamba_kernels=False (#44853) by @sergiopaniego in [#44853]
• revert sha commit pointing to main for transformers_amd_ci_ workflows (#45495) by @paulinebm in [#45495]
• Fix ZeRO-3 from_pretrained: load registered buffers in _load_state_dict_into_zero3_model (#45402) by @saslifat-gif in [#45402]
• Remove redundant condition checks in get_image_size method (#45461) by @JiauZhang in [#45461]
• Add check-auto in repo-consistency and fix sorting (#45481) by @zucchini-nlp in [#45481]
• Fix typos in src/transformers/utils/output_capturing.py (#45269) by @ryota-komatsu in [#45269]
• typing: rule 15 - checks for tie_word_embeddings presence (#44988) by @tarekziade in [#44988]
• [CB] Fix capture of max_seqlen (#45323) by @remi-or in [#45323]
• Minor update (#45484) by @ydshieh in [#45484]
• Add Neuron to auto-compile hardware list (#44757) by @dacorvo in [#44757]
• Allow loading Qwen Thinker 'base' models without generative head (#45457) by @tomaarsen in [#45457]
• [fix] Always early return for non-Mistral models in _patch_mistral_regex (#45444) by @tomaarsen in [#45444]
• Fix spurious position_ids warnings for at least 40 architectures (#45437) by @tomaarsen in [#45437]
• [fix] Make Qwen2_5OmniProcessor warning a lot less noisy via warning_once (#45455) by @tomaarsen in [#45455]
• Dynamic auto mapping (#45018) by @zucchini-nlp in [#45018]
• [docs] vlm addition (#45271) by @stevhliu in [#45271]
• fix: dont download artifacts from the test hub (#45319) by @tarekziade in [#45319]
• fix(clipseg): fix 2 failing tests (#45403) by @kaixuanliu in [#45403]
• [docs] @auto_docstring decorator (#45130) by @stevhliu in [#45130]
• Fix Sam3Processor missing input_boxes_labels for padded None entries (#45171) by @Kash6 in [#45171]
• better grad acc tests (#45434) by @SunMarc in [#45434]
• Add example for iterative chatting with MLLMs (#45398) by @zucchini-nlp in [#45398]
• Gemma4 resizing per layer inputs (#45324) by @zucchini-nlp in [#45324]
• Add step3_vl to MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS (#45449) by @hmellor in [#45449]
• Update workflow references to new commit hash (#45442) by @paulinebm in [#45442]
• [Gemma4] Add docstrings for Per-Layer Embeddings (PLE) pipeline (#45207) by @w4nderlust in [#45207]
• [Doc] Correct checkpoint path in Dinov2 model_docs (#45430) by @ambroiseodt in [#45430]
• Fix ty for transformers cli (#45190) by @SunMarc in [#45190]
• fix(models): Resolve regressions in Wav2Vec2PhonemeCTCTokenizer (wav2vec2-lv-60-espeak-cv-ft) (#45199) by @harshaljanjani in [#45199]
• Fix Qwen2.5VL temporal grid positions (#45400) by @zucchini-nlp in [#45400]
• [fix] PEFT integration fixes preventing save/load & integration (#45428) by @tomaarsen in [#45428]
• Fix the response schema for the gemma4 converter (#45411) by @Rocketknight1 in [#45411]
• [Doc] MoE routing capture and replay recipe (#44925) by @kashif in [#44925]
• Fix apply_chat_template crash on tool_call messages without content (#45348) by @qgallouedec in [#45348]
• [AMD CI] Fix torch.compile/export failures on AMD CI due to untraceable set.contains (#45282) by @Abdennacer-Badaoui in [#45282]
• [inference_fusion] convert conv3d patch embed to linear (#45041) by @JJJYmmm in [#45041]
• Fix #45305 + add regression test GAS (#45349) by @florian6973 in [#45349]
• Update trackio integration to use Buckets and "freeze" Space after training (#45329) by @abidlabs in [#45329]
• fix(qwen3_moe): correct return type annotation on Qwen3MoeSparseMoeBlock.forward (#45352) by @RudrenduPaul in [#45352]
• Fix: NotebookProgressCallback crash when evaluating with the Trainer (#44949) by @Charly21r in [#44949]
• docs: fix 5 docstring errors in Gemma3nTextConfig (typos, grammar, formatting) (#45370) by @RudrenduPaul in [#45370]
• Less unnecessary RoPE warnings (#45289) by @zucchini-nlp in [#45289]
• Fix unintended Hub metadata calls from _patch_mistral_regex (#43603) by @vaibhav-research in [#43603]
• Fix MoE routers returning probabilities instead of logits (#45131) by @yacinemebarki in [#45131]
• [docs] training on specific hardware (#44799) by @stevhliu in [#44799]
• [docs] zero + sequence parallelism (#44605) by @stevhliu in [#44605]
• Fix vlm weight mappings (#45358) by @Cyrilvallez in [#45358]
• Copy the template resolution logic from the base apply_chat_template to Voxtral (#45117) by @Rocketknight1 in [#45117]
• add kwargs to all methods in the CallbackHandler class (#45353) by @wilnn in [#45353]
• Close file handler (#45187) by @ydshieh in [#45187]
• fix: restore mypy type checking for PreTrainedConfig subclasses (#45071) (#45240) by @shhKnight30 in [#45240]
• cohere_asr: fix device issue for test_model_parallel_beam_search (#45214) by @kaixuanliu in [#45214]
• Fix AttributeError in Gemma3ForConditionalGeneration and Gemma3ForSequenceClassification when config.return_dict=False (#45277) by @kamalrajkannan78 in [#45277]
• fix bug for videomt model device mismatch (#45204) by @kaixuanliu in [#45204]
• fix gemma4 gradient accumulation loss and last token incorrect labels (#45354) by @winglian in [#45354]
• Logger has [transformers] prefix in non-verbose mode (#45316) by @zucchini-nlp in [#45316]
• Fix AttributeError in AssistantToTargetTranslator.unmap_input_ids with cross-vocab models (#45320) by @Regata3010 in [#45320]
• musicflamingo: add test support for Intel XPU device (#45212) by @kaixuanliu in [#45212]
• nomic_bert: make the test suitable for general device. (#45209) by @kaixuanliu in [#45209]
• Skip invalid flash-attn tests for pi0 model (#45011) by @kaixuanliu in [#45011]
• Add cuda compatibility check for using grouped_mm (#45001) by @Sai-Suraj-27 in [#45001]
• [docs] optimizers, hyperparam search, training features (#44290) by @stevhliu in [#44290]
• Remove unused parameters and improve add_tensor_parallel_hooks_t… (#44768) by @michaelbenayoun in [#44768]
• [gemma4] Fix device map auto (#45347) by @Cyrilvallez in [#45347]
• Refactor CLIP-like models (#44431) by @zucchini-nlp in [#44431]
• refactor: display test duration (#45344) by @tarekziade in [#45344]
• Fix Wav2Vec2Config.vocab_size type to allow None (#45108) by @jiqing-feng in [#45108]
• Add THD support in ESM (#44145) by @balvisio in [#44145]
• [gemma4] Remove all shared weights, and silently skip them during loading (#45336) by @Cyrilvallez in [#45336]
• Fix conversion mappings for vlms (#45340) by @Cyrilvallez in [#45340]
• chore: added circleci python script to ruff and ty checkers (#45339) by @tarekziade in [#45339]
• tweak checkers output on errors (#45163) by @tarekziade in [#45163]
• chore: remove test_hub for now (#45337) by @tarekziade in [#45337]
• [docs] pipeline cleanup (#44954) by @stevhliu in [#44954]
• Fix export for gemma4 and add Integration tests (#45285) by @Cyrilvallez in [#45285]
• Fix vllm cis (#45139) by @ArthurZucker in [#45139]
• [docs] static model rules (#45232) by @stevhliu in [#45232]
• fix(security): prevent untrusted users from triggering TRL CI dispatch (#45302) by @jagwar in [#45302]
• [AMD CI] Fix Qwen2 expectations (#45284) by @Abdennacer-Badaoui in [#45284]
• Add hasattr(torch.backends.cudnn, "conv") to conftest.py (#45263) by @ydshieh in [#45263]
• Fix SmolVLM video processor resize using wrong interpolation after backend refactor (#45258) by @ydshieh in [#45258]
• Fix Qwen2IntegrationTest (#45268) by @ydshieh in [#45268]
• doc: fix TokenizersBackend.convert_to_native_format docstring (#45262) by @lowzhao in [#45262]
• empty (#45261) by @ydshieh in [#45261]
• Fix unexpected TF32 being enabled in testing (#45252) by @ydshieh in [#45252]
• Fix tf32 issue: set torch.backends.cudnn.conv.fp32_precision explicitly. (#45248) by @ydshieh in [#45248]
• Nvidia CI with torch 2.11 (#45243) by @ydshieh in [#45243]
• Update tiny model creation script (#45241) by @ydshieh in [#45241]
• Update get_test_info.py (related to tiny model creation) (#45238) by @ydshieh in [#45238]
• More fix for tiny model creation (#45228) by @ydshieh in [#45228]
• remove unnecessary entries in some auto model mappings (#45224) by @ydshieh in [#45224]
• fix: hf-doc-builder insallation was failing (#45225) by @tarekziade in [#45225]
• [CB] Add per-request logits processors (#45026) by @remi-or in [#45026]
• [docs] formatting (#45196) by @stevhliu in [#45196]
• fix test_register_result_handler (#45188) by @SunMarc in [#45188]
• [CB] Tweaks to update and minor fixes (#45179) by @remi-or in [#45179]
• Fix pypi release (#45210) by @ArthurZucker in [#45210]
• fix(docs): correct gemma4 docs and examples (#45197) by @douglas-reid in [#45197]
• Add Turkish (tr) translation for Get Started section (#45158) by @onwp in [#45158]

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @vasqu
  • [Privacy Filter] Add model (#45580)
  • Fix typos (#45574)
  • [Conversion Mapping] Small fixups (#45483)
  • 🚨 [Kernels] Fix kernel function registration (#45420)
  • [Tokenizers] Move gpt sw3 tokenizer out (#45404)
• @rain-1
  • Add /v1/completions endpoint (OpenAI legacy completions API) to transformers serve (#44558)
• @zhang-prog
  • Updated the image cache for Paddle models according to the latest API (#45562)
  • [Model] Add SLANet Model Support (#45532)
  • Fix resize failure caused by zero-sized masks in PP-DocLayoutV3 (#45281)
• @tarekziade
  • fix table update versions (#45544)
  • qa: re-run modular converter when the script itself is modified (#45528)
  • Revert "Fix: modular image processors (#45492)" (#45531)
  • chore(qa): split out mlinter (#45475)
  • typing: rule 15 - checks for tie_word_embeddings presence (#44988)
  • fix: dont download artifacts from the test hub (#45319)
  • refactor(qa): extend extras so ty can run on server modules (#45456)
  • remove cache file from tree (#45392)
  • refactor: display test duration (#45344)
  • http retries on audio file downloads (#45126)
  • chore: added circleci python script to ruff and ty checkers (#45339)
  • tweak checkers output on errors (#45163)
  • fix: leak in tokenizer registry for test_processors (#45318)
  • chore: remove test_hub for now (#45337)
  • fix: hf-doc-builder insallation was failing (#45225)
• @marvinzh
  • add Qianfan-OCR model definition (#45280)
• @remi-or
  • [CB] Fix capture of max_seqlen (#45323)
  • [CB] Add per-request logits processors (#45026)
  • [CB] Tweaks to update and minor fixes (#45179)
• @ydshieh
  • Minor update (#45484)
  • Close file handler (#45187)
  • Add hasattr(torch.backends.cudnn, "conv") to conftest.py (#45263)
  • Fix SmolVLM video processor resize using wrong interpolation after backend refactor (#45258)
  • Fix Qwen2IntegrationTest (#45268)
  • empty (#45261)
  • Fix unexpected TF32 being enabled in testing (#45252)
  • Fix tf32 issue: set torch.backends.cudnn.conv.fp32_precision explicitly. (#45248)
  • Nvidia CI with torch 2.11 (#45243)
  • Update tiny model creation script (#45241)
  • Update get_test_info.py (related to tiny model creation) (#45238)
  • More fix for tiny model creation (#45228)
  • remove unnecessary entries in some auto model mappings (#45224)
• @NielsRogge
  • Add SAM3-LiteText (#44320)
• @ArthurZucker
  • Fix IndexError with DeepSpeed ZeRO-3 when kernels rotary is active (#45414)
  • Fix Kimi-K2.5 tokenizer regression and _patch_mistral_regex AttributeError (#45359)
  • Fix vllm cis (#45139)
  • Fix pypi release (#45210)
  • update to dev version 5.6.0-dev0
• @JJJYmmm
  • [inference_fusion] convert conv3d patch embed to linear (#45041)
• @balvisio
  • Add THD support in ESM (#44145)
• @onwp
  • Add Turkish (tr) translation for Get Started section (#45158)

This is mostly some fixes that are good to have asap, mostly for tokenizers; ** Fix Kimi-K2.5 tokenizer regression and _patch_mistral_regex Attribute… (#45305) by ArthurZucker

For training: ** Fix #45305 + add regression test GAS (#45349) by florian6973, SunMarc ** Fix IndexError with DeepSpeed ZeRO-3 when kernels rotary is active (#…) by ArthurZucker

And for Qwen2.5-VL : ** Fix Qwen2.5-VL temporal RoPE scaling applied to still images (#45330) by Kash6, zucchini-nlp

<img width="2786" height="1504" alt="image" src="https://github.com/user-attachments/assets/6c8c878f-042b-4858-9f64-73fd9ccd7e4b" />

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

<img width="1478" height="1374" alt="image" src="https://github.com/user-attachments/assets/9d88bd1b-02ea-4829-b7d0-fac0e347d436" />

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @lashahub in #43538

Breaking changes

Mamba and hybrid model caches are now first-class native citizens in the library, so users working with Mamba-based or hybrid (Mamba + attention) models should update their code to use the new native cache classes instead of any previous workarounds.

• 🚨 [Cache] Native mamba & hybrid cache (#44950) by @Cyrilvallez

Remote code execution support has been removed from the native LightGlue integration, so users who were loading LightGlue with trust_remote_code=True must remove that argument and use the model directly through the standard native API.

• 🚨 [LightGlue] Remove remote code execution (#45122) by @vasqu

Vision

Several vision-related bugs were fixed in this release, including correcting the Gemma vision mask to support video inputs, resolving a dependency issue that incorrectly required torchvision for PIL-based image processors, and patching bugs in the Janus image generation model and image loading. Local code resolution for tokenizers and image processors was also corrected.

• Generalize gemma vision mask to videos (#45185) by @zucchini-nlp in [#45185]
• Fix explicit local code resolution for tokenizers and image processors (#45169) by @hmellor in [#45169]
• fix bug for janus model image generation (#45044) by @kaixuanliu in [#45044]
• [Bugfix] Remove incorrect torchvision requirement from PIL backend image processors (#45045) by @Lidang-Jiang in [#45045]
• Avoid Image.open failure (#44645) by @sywangyi in [#44645]

Cache

Improved the performance of repository checks (check-repo) by introducing file-level and AST-level disk caching, achieving up to a 27x speedup (from ~46s to ~1.6s with a warm cache), and fixed the mlinter cache location in .gitignore.

• refactoring: speedup static checks with disk cache (#44992) by @tarekziade in [#44992]
• refactor: added cache in check_repo (#45012) by @tarekziade in [#45012]
• chore: Fix mlinter cache location (#45052) by @tarekziade in [#45052]

Bugfixes and improvements

• Fix resized LM head weights being overwritten by post_init (#45079) by @javierdejesusda in [#45079]
• [Qwen3.5 MoE] Add _tp_plan to ForConditionalGeneration (#45124) by @danielquintas8 in [#45124]
• fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#45074) by @harshaljanjani in [#45074]
• [misc] fix qwen35 tests: correct the text model type and skip reverse_mapping (#45173) by @JJJYmmm in [#45173]
• 🔒 Pin GitHub Actions to commit SHAs (#45180) by @paulinebm in [#45180]
• Use doc-builder runnable example for GLM-ASR (#44277) by @tarekziade in [#44277]
• CI] Small T5 expectations updated (#45138) by @Abdennacer-Badaoui in [#45138]
• fix: correct type annotations across config classes for @strict validation (#45007) by @Krishnachaitanyakc in [#45007]
• Fix T5Attention shape mismatch under Tensor Parallelism (#45109) by @aws-zhanxun in [#45109]
• [refactor] Serving into proper modules (#44796) by @SunMarc in [#44796]
• Re-add regex substitutions to the response parsing spec (#45166) by @Rocketknight1 in [#45166]
• Fix incorrect TrainingArguments example in training.md (#45150) by @maanas1234 in [#45150]
• Add parse_response to Processor, make it a bit more official (#45143) by @Rocketknight1 in [#45143]
• DeepGEMM (#44832) by @IlyasMoutawwakil in [#44832]
• fix: prefer registered config over remote code in AutoConfig.from_pretrained (#45094) by @HanFa in [#45094]
• [serving] Fix continuous batching JSON response serialization (#45057) by @NathanHB in [#45057]
• Fix stupid test fetcher (#45140) by @ydshieh in [#45140]
• [CB] Add warmup feature (#45112) by @remi-or in [#45112]
• feature: added import complexity checker (#45013) by @tarekziade in [#45013]
• Fix tests for janus model (#44739) by @kaixuanliu in [#44739]
• CB improvements for serving (#45063) by @SunMarc in [#45063]
• [docs] continuous batching (#44896) by @stevhliu in [#44896]
• Fix few issues in Qwen_3_Omni_Moe (#44848) by @Sai-Suraj-27 in [#44848]
• Fix TypeError in rope validation when ignore_keys is a list (#45069) by @Fr0do in [#45069]
• Remove unused TensorFlow env var (#45065) by @Sai-Suraj-27 in [#45065]
• fix: add identity reverse_op to dequantize ops for save_pretrained (#44983) by @Hyungkeun-Park-Nota in [#44983]
• Fix when RoPE params are in kwargs (#45049) by @zucchini-nlp in [#45049]
• chore: update update_metdata.yml (#45054) by @hf-security-analysis[bot] in [#45054]
• [FA] Fix BC support for a few versions + add deprecation cycle (#45061) by @vasqu in [#45061]
• fix(testing): Fix Parakeet, Evolla, Pi0, and Phi-3 test failures on main CI (#45004) by @harshaljanjani in [#45004]
• Allow advanced users to override model_type in AutoConfig.from_pretrained (#45058) by @hmellor in [#45058]
• Fix failing SmolLM3IntegrationTest (#45048) by @Sai-Suraj-27 in [#45048]
• chore: remove old extras (#45024) by @tarekziade in [#45024]
• Embedding VLMs don't need a head (#45000) by @zucchini-nlp in [#45000]
• Fix GraniteConfig type hints to accept int for multiplier fields (#45019) by @javierdejesusda in [#45019]
• fix: preserve rotary_pct across save/load cycle in GPTNeoX configs (#44985) by @Krishnachaitanyakc in [#44985]

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @ed22699
  • Internalise the NomicBERT model (#43067)
• @tarekziade
  • Use doc-builder runnable example for GLM-ASR (#44277)
  • refactoring: speedup static checks with disk cache (#44992)
  • feature: added import complexity checker (#45013)
  • refactor: added cache in check_repo (#45012)
  • chore: remove old extras (#45024)
  • chore: Fix mlinter cache location (#45052)
  • refactor: speed up docstring checker (#45009)
• @Krishnachaitanyakc
  • fix: correct type annotations across config classes for @strict validation (#45007)
  • fix: preserve rotary_pct across save/load cycle in GPTNeoX configs (#44985)
• @lashahub
  • Add Music Flamingo (#43538)
• @Lidang-Jiang
  • [Bugfix] Remove incorrect torchvision requirement from PIL backend image processors (#45045)