Hugging Face

Transformers v5.13 added Kimi K2.5 support, Diffusers landed Cosmos 3, and huggingface_hub shipped Sandboxes.

New model support in Transformers and Diffusers. Transformers v5.13 added the Kimi K2.5 architecture (used by K2.5–2.7), an open-source multimodal agentic model. Diffusers v0.39.0 added NVIDIA's unified Cosmos 3 world-model pipeline, spanning generation, reasoning, and action conditioning.

Sandboxes: isolated cloud machines from the CLI. huggingface_hub v1.22 shipped Sandboxes — ephemeral VMs spun up in seconds, with live streaming, file transfer, and port proxying. Two flavors exist: dedicated Sandbox.create (GPU workloads, full isolation) and SandboxPool for packing CPU sandboxes into shared hosts (RL rollouts). Background processes and port proxy URLs are supported.

TRL v1.8 graduated KTO and added environment-owned rewards. KTO is now a stable trainer at the top-level trl package (experimental path still works with a FutureWarning). Agentic RL training gained environment-owned rewards, multi-environment support, and a new response parsing API.

huggingface_hub v1.23 added Space templates and smoother Xet downloads. Seed a new Space from official templates (JupyterLab, Gradio chatbot, Streamlit) via list_space_templates() or hf spaces templates. The Space SDK is inferred from the template. CLI extensions can be added via hf extensions install, and Xet downloads are more reliable.

LeRobot v0.6 overhauled its RL API with breaking changes. The RL stack was rebuilt: the sac policy type is now gaussian_actor under a new modular API. pip install lerobot no longer includes dataset or training dependencies (use extras). GR00T N1.5 is replaced by N1.7. Minimum PyTorch is now 2.7. eval_freq renamed to env_eval_freq. Legacy subtask_index annotations replaced by language columns.

timm v1.0.28 added EUPE ViT, TIPSv2 models, and optimizer torch.compile support. New model definitions and pretrained weights for Efficient Universal Perception Encoder (EUPE) and TIPSv2 vision encoders. Optimizer torch.compile and tensor learning-rate support. NaFlexViT patch-layout extended.

Routine patches. Transformers v5.13.1 fixed vLLM sync. TRL v1.7.1 fixed GRPO+vLLM+PEFT hangs and dataset fingerprinting.