Hugging Face

Jobs on the Hub now support an optional --name flag on the CLI and a name parameter on the Python API, making them easier to find in the UI. The README was recast to put the hf CLI first, with a new "For AI agents" section.

Fixed assisted decoding for models using EncoderDecoderCache (e.g., OlmoHybrid) and a SDPA prefill issue with position_bias during StaticCache. Also bumps FP8 kernels and fixes DeepGEMM on multiple devices.

Added Inkling, a 975B-parameter multimodal model, and TIPSv2. GPTNeoX now remaps embed_out to lm_head and GPTBigCode has attention backend changes for vLLM compatibility. Multi-Token Prediction decoding support, SDPA prefill with FlashAttention for StaticCache (up to 260% faster), and numerous bug fixes across MoE, cache, and generation.

Hardened pickle loading against security concerns, fixed a ZeroDivisionError in AverageMeter, and fixed reg_token not being added to no-weight-decay and stem/embed LR groups. Also added EUPE ViT (DINOv3-style) and TIPSv2 (DINOv2-style) model definitions with pretrained weights, improved optimizer torch.compile support, and extended NaFlexViT patch-layout and forward_intermediates support.

Fixes custom model compatibility with the latest vllm release by being more defensive with remap_legacy_layer_types and handling cases where custom code doesn't know about the new linear layer type names. Also fixed a key type assertion in _LazyAutoMapping.register.

KTO trainer graduates from experimental to the top-level trl package with the same API as DPO/GRPO/SFT, and the experimental import path still works with a FutureWarning. Environment-owned rewards let agentic RL environments define their own reward via a reserved get_reward() method, and multi-environment support allows a single training run to handle multiple environments with environment-specific tool schemas. GRPO now supports both static and adaptive entropy regularization to encourage exploration and prevent policy collapse.

LeRobot v0.6.0 makes dataset and training dependencies optional, installed only when needed (e.g., pip install lerobot[training]). The RL stack was rebuilt with a new modular API: the sac policy type is now gaussian_actor. GR00T N1.5 is replaced by N1.7, PyTorch minimum is now 2.7, import paths have been consolidated to canonical public entry points, eval_freq renamed to env_eval_freq, legacy subtask_index annotations replaced by language columns, and video codec configuration split per camera. Also added five new benchmark integrations (RoboCasa365, RoboTwin 2.0, RoboCerebra, RoboMME, LIBERO-plus), 2x faster dataloader via parallel decode, and a new lerobot-rollout CLI to decouple policy deployment from data recording.

Fixed a hang in GRPO + vLLM colocate + PEFT on non-NVLink hardware and corrected dataset fingerprinting in DPO/SFT tokenization. Also integrated the new response parsing API, added a prompt-learning guard for PEFT with Liger in GRPO, and fixed activation offload storage deduplication.

This release adds support for seven new model architectures: Kimi 2.5–2.7 (multimodal agentic coding), MiMo-V2-Flash (256K context MoE model), Nemotron 3.5 ASR and Nemotron ASR Streaming (multilingual speech recognition with configurable latency-accuracy tradeoffs), Qwen3 ASR with forced aligner, ZAYA1 (MoE language model), VideoPrism (video understanding encoder), and RADIO (vision foundation model family).

Sandboxes introduce isolated cloud VMs with live-streamed output, file transfer, and a proxy for reaching in-sandbox servers. Snapshot downloads now cache repo file listings on disk for a single-network-call re-download. The CLI was rebuilt on Click, deprecating Typer. Breaking changes include case-sensitive pattern matching on all platforms, removal of six inference providers, and deprecation of upload_large_folder.

Nine new pipelines are now available: Cosmos 3 (NVIDIA's unified world foundation model combining generation, reasoning, and action), Ideogram 4 (flow-matching text-to-image with asymmetric guidance), Krea 2 (single-stream MMDiT with Qwen3-VL encoding), DreamLite (ByteDance text-to-image and editing), PRX Pixel (pixel-space generation by Photoroom), Motif-Video (2B text-to-video and image-to-video), AnyFlow (any-step video diffusion from NVIDIA/NUS/MIT), JoyAI-Image-Edit (unified multimodal for editing with spatial control), and DiffusionGemma (block-diffusion language model). Additional enhancements across schedulers, training utilities, and core library components.

The default SFT loss_type is now "chunked_nll", delivering ~30% less peak VRAM on average with neutral or slightly faster wall-clock time. Also introduces experimental GMPO trainer, transformers continuous batching, AsyncGRPO weight sync with vLLM 0.22+, and paddding-free AsyncGRPO.

Browser-based OAuth Device Code flow is now the default login method, and hf upload uses a streamed multi-commit pipeline for faster, resumable large folder uploads. Also adds job wait commands, SSH access to running Jobs, and custom-container deployment for Inference Endpoints, with breaking changes around multi-commit uploads and non-detached job exit codes.

Fixed mistral tokenizer resolution when mistral-common is installed and updated the lower bound for PEFT. This is similar to v5.10.3 minus fixes already in the main release.