Hugging Face

The tensor viewer now sends a single request for the headers of all shards instead of fetching each shard one by one, cutting load times on large sharded models such as DeepSeek-V3.2 (9.2s to 0.45s), GLM-5.3 (4.0s to 0.57s), and Qwen2.5-72B-Instruct (1.2s to 0.27s).

Fine-grained access tokens and service account tokens can now be granted a contributor permission that allows creating repositories in an account or selected organizations, with full write access limited to repositories the token created. It mirrors the contribute-repos scope for OAuth apps and targets CI/CD pipelines, training jobs, and agents publishing checkpoints or datasets.

Fixes a breaking truncation and overflowing regression introduced in v0.23.1, and loosens the huggingface_hub requirement to allow v2, unblocking transformers installs. Also fixes type hints not resolving for pyright.

Decision models now have a dedicated decision-model tag, an icon on model pages, and a filter in the Misc section of the Models page. GGUF decision models are tagged automatically from their llama.cpp metadata.

A new fused LM head, on by default in SFT, DPO, KTO, GRPO, RLOO and Distillation, projects hidden states through the LM head in Triton-kernel tiles to avoid building the full logits tensor, raising max trainable sequence length up to 6.9x and cutting peak memory 52% to 82%. The full-logits scoring path, use_liger_kernel chunked path and _forward_redirection are removed, so a PEFT adapter on lm_head now raises and use_liger_kernel=True is deprecated for v2.0.0 removal.

hf jobs stats and HfApi.fetch_job_metrics now return a single snapshot instead of streaming until job completion, with -f/--follow on the CLI and follow=True in Python restoring the previous behavior. Also fixed a race in _get_version that made concurrent downloads skip hf_xet, and multiple HfFileSystem cache issues including duplicate listings and stale not-found results.

Adds support for EmbeddingGemma 2, an open multimodal embedding model that maps text, images, video, and audio into a single unified 768-dimensional vector space. Text search works via the feature-extraction pipeline, while other modalities use AutoProcessor and AutoModel directly, with all inputs comparable through cosine similarity.

GRPO, RLOO, and Distillation on the default non-vLLM generation path now stop completions at every eos id the model declares, fixing silent training on tokens generated past the end of the turn that showed up as completions/clipped_ratio near 1. Also fixes DFT loss on batches with no trainable tokens, precompute_ref_log_probs under FSDP with no reference model, completion offsets where the tokenized prompt and prompt+completion diverge, and guards the requests and urllib3 imports in the vLLM client.

Qwen-Image 2.1 unifies image generation and editing in one model with 7B visual generation parameters, multiple reference images, native RGBA transparency output, and LoRA training. ONNX support is deprecated in favor of Optimum, and previously deprecated LuminaText2ImgPipeline and Lumina2Text2ImgPipeline aliases were removed in favor of LuminaPipeline and Lumina2Pipeline.

Datasets 5.1.0 adds the Vortex columnar format, Harbor RL environments, and biological formats including FASTA, FASTQ, GenBank, PDB, and mmCIF with BioSequence/BioStructure decoding. Fixes include an archive path check that allowed sibling-directory escapes, disallowed HDF5 external links, and an infinite-repetition bug in empty iterable datasets.

Users can now add a P(doom) estimate, their probability of AI causing an existential catastrophe, to their Hugging Face profile. Answers also feed a community survey on AI risk perception, with only aggregated, anonymized results published in two weeks.

Fixes an issue that prevented encoder-decoder models from working when using Transformers 5.18.0 or later.