Hugging Face

Recently shipped Transformers v5.16 with GLM-5.3-Flash and Qwen4-Exp, plus a TRL rewrite that delegates serving to vLLM's own server.

Transformers v5.16 added two flagship models. GLM-5.3-Flash is the first natively multimodal GLM-5 model: 320B total parameters, 18B active, with a hybrid sparse-and-linear attention architecture that cuts long-context serving costs. Qwen4-Exp brings GatedResidual, Qwen Sparse Attention, and Per-Layer Embedding to the Qwen line. Patch v5.16.1 also included GLM support. [source: https://github.com/huggingface/transformers/releases/tag/v5.16.0] [source: https://github.com/huggingface/transformers/releases/tag/v5.16.1]

TRL's vLLM serving now delegates to vLLM's own server. The custom FastAPI wrapper dropped from ~1,218 to ~130 lines, and weight sync uses vLLM's NCCL engine — one packed broadcast instead of per-tensor HTTP requests. fastapi, uvicorn, and pydantic left the vllm extra. Note: v1.12.0 is a bit-identical accidental duplicate of v1.11.0; the next real release is v1.13.0. [source: https://github.com/huggingface/trl/releases/tag/v1.11.0] [source: https://github.com/huggingface/trl/releases/tag/v1.12.0]

Diffusers v0.40.0 graduated Modular Diffusers out of the experimental phase. The release added LTX2.5, Wan Animate 2, and MiniMax-H3, which generates video and its soundtrack together in a single transformer pass — no separate vocoder. It also introduced minimal tensor-parallel support. [source: https://github.com/huggingface/diffusers/releases/tag/v0.40.0]

huggingface_hub v1.29.0 fixed Xet download rate limits. Large repos no longer make one API call per file — a 77k-file repo dropped from ~1,500 Hub API calls per minute to a single cached connection. Bucket visibility can now be changed after creation. [source: https://github.com/huggingface/huggingface_hub/releases/tag/v1.29.0]

timm v1.0.29 added LowFormer, CPUBone, PP-LCNetV2, and LingBot-Vision with pretrained weights, plus per-batch image and batch-size scheduling for progressive resolution training. [source: https://github.com/huggingface/pytorch-image-models/releases/tag/v1.0.29]

LeRobot renamed its types module. lerobot.types moved to lerobot.lerobot_types in v0.6.1 — a breaking change for code importing constants from the old path. [source: https://github.com/huggingface/lerobot/releases/tag/v0.6.1]