Who contributed to Hugging Face on August 27, 2026?

12 developers shipped this update, including Cyrilvallez, Steven Liu, dkrisman, guarin, Harry Mellor, shaun smith, osolmaz, and tarekziade, and 4 more.

What were the notable Hugging Face updates?

Release: v5.16.0, [docs] Fix links and remove TokenizerFast (#47748), and Add an opt-in per-frame pixel cap (cap_pixels_per_frame) to the Qwen3-VL video processor.

Transformers v5.16.0 ships with qwen4-exp · Hugging Face

Hugging Face's flagship release lands with a brand-new Qwen architecture that changes how residuals and sparse attention are built.

Transformers v5.16.0 lands with Qwen4-Exp [1], a hybrid model that marries GatedResidual and Qwen Sparse Attention, and it's the kind of release that resets the baseline for the whole stack. The attention mechanism [1] "reduces per-layer KV cache storage requirements" [1], which should make long-context inference materially cheaper for anyone moving to the new checkpoint. While you're upgrading, note that the docs cleanup [2] drops TokenizerFast references, and the video processor now caps pixels per frame [3] for Qwen3-VL, both worth pulling in before your next build. Harbor-HF spent the day stabilizing its agent runner: a Docker Hub quota blowout [4] where four tasks burned both anonymous pull attempts is fixed by collapsing each task image to a single remote copy, and two worker pins [5][6] lock that fix and the leading-hyphen repair [7] into every Terminal-Bench profile. On the tooling side, the ML linter now stops nagging about config plumbing [8] and single-source MoE widths [9], so deepseek and qwen2_moe PRs will stop tripping false positives, and the hub client caches Xet connection info [10] to kill the one-token-request-per-file storm that was rate-limiting 77k-file snapshot downloads. Finally, the CI health dashboard was crying wolf over 2.39M "lost" spans [11] that the privacy filter actually dropped on purpose, so the red alert you saw yesterday was noise.

@huggingface

TRANSFORMERS V5.16.0 SHIPS WITH QWEN4-EXP

TRL UNBLOCKS MOE LORA AS PEFT PATCHES AUTOCAST

Transformers MoE loss cut 99.7% as TRL gains tool calling

Transformers linter hunts nn.Buffer misuse as kernels ships hotfix

Harbor-hf sinks broken Sandbox cleanup

Cline desktop plugins crash on launch

vLLM 0.28.0 LANDED WITH 584 COMMITS AND A KIMI-K3 PERFORMANCE PUSH

CachyOS Tuned Strikes: THP and Laptop Mode Out

ECT0 FIXES WARNINGS-AS-ERRORS TYPO