Who contributed to Hugging Face on July 25, 2026?

9 developers shipped this update, including Matt, Cyril Vallez, Anton Vlasjuk, qgallouedec, albertvillanova, AmineDiro, Steven Palma, and CarolinePascal, and 1 more.

What were the notable Hugging Face updates?

Deprecate the old response_schema (#47320), Refactor all linear attention models to latest best standards for convolution (#47452), and Add AXK2 from SKT (#47528).

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

TRANSFORMERS OVERHAULS LINEAR ATTENTION WHILE DEPRECATING LEGACY RESPONSE SCHEMA

By RepoJournal · Filed 06:59 UTC on July 25, 2026 · About Hugging Face

9 people shipped this

qgallouedec @qgallouedec 3 cited

Steven Palma @imstevenpmwork 2 cited

suvadityamuk @suvadityamuk 2 cited

Matt @Rocketknight1 1 cited

Cyril Vallez @Cyrilvallez 1 cited

Anton Vlasjuk @vasqu 1 cited

AmineDiro @AmineDiro 1 cited

CarolinePascal @CarolinePascal 1 cited

albertvillanova @albertvillanova

The transformers library is retiring its fragile response_schema prototype in favor of streaming-compatible parsing, while simultaneously refactoring every linear attention model to use standardized convolution patterns.

Transformers dropped the old response_schema [1] after discovering it was brittle, hard to write regexes for, and couldn't handle streaming. The deprecation triggers a cascade: TRL now ships a new Gemma4 response template [4] to avoid breaking on models that relied on the old approach. Simultaneously, the library refactored all linear attention models to latest convolution best practices [2], touching Mamba, Jamba, Falcon, Qwen, Zamba and others in a massive consolidation that simplifies cache handling and prefill logic. SKT's AXK2 landed as a new flagship LLM [3], refactored to inherit from DeepSeek V3 architecture. On the TRL side, AsyncGRPOTrainer is gaining a loop-owning agent path [5] for training external tools that manage their own loops like OpenCode, plus DistillationTrainer continues its incremental refactor with new generation methods [6] [7]. LeRobot bumped diffusers from 0.36 to 0.38+ [8] to pick up a critical security fix blocking downstream consumers, and added token argument support for private Hub datasets [9] while fixing depth frame handling in image preprocessing [10]. Skills expanded Spaces templates to cover 3D generation models [11], then refined the documentation to avoid skewing agent behavior [12].

Action items

→ Review deprecated response_schema usage in any custom parsing code and migrate to streaming-compatible solution huggingface/transformers [plan]
→ Upgrade diffusers to >=0.38.0 if consuming lerobot or dependent on security fix GHSA-98h9-4798-4q5v huggingface/lerobot [plan]
→ Test linear attention model outputs if running inference on Mamba, Jamba, Qwen, Zamba or Falcon variants huggingface/transformers [monitor]

References

[1] Deprecate the old response_schema (#47320) huggingface/transformers
[2] Refactor all linear attention models to latest best standards for convolution (#47452) huggingface/transformers
[3] Add AXK2 from SKT (#47528) huggingface/transformers
[4] Add Gemma4 response template ↗ huggingface/trl
[5] Async grpo OpenEnv harness rollout ↗ huggingface/trl
[6] [DistillationTrainer refactor] Add `_generate_single_turn` (unwired) (#6513) huggingface/trl
[7] [DistillationTrainer refactor] Add `_generate` + `_generate_and_score_completions` (unwired) (#6521) huggingface/trl
[8] chore(deps): bump diffusers to `>=0.38.0,<0.40.0` (#4145) huggingface/lerobot
[9] feat(dataset): accept token argument for private HF Hub datasets ↗ huggingface/lerobot
[10] (depth image processing): excluding depth frames from the RGB to BGR image processing ↗ huggingface/lerobot
[11] huggingface-spaces: add 3D generation model coverage within the Spaces skill ↗ huggingface/skills
[12] huggingface-spaces: Remove disproportionate 3D generation information ↗ huggingface/skills

Quick answers

What shipped in Hugging Face on July 25, 2026?: The transformers library is retiring its fragile response_schema prototype in favor of streaming-compatible parsing, while simultaneously refactoring every linear attention model to use standardized convolution patterns. In total, 35 commits and 35 pull requests landed.
Who contributed to Hugging Face on July 25, 2026?: 9 developers shipped this update, including Matt, Cyril Vallez, Anton Vlasjuk, qgallouedec, albertvillanova, AmineDiro, Steven Palma, and CarolinePascal, and 1 more.
What were the notable Hugging Face updates?: Deprecate the old response_schema (#47320), Refactor all linear attention models to latest best standards for convolution (#47452), and Add AXK2 from SKT (#47528).

TRANSFORMERS SHIPS FSDP DISTRIBUTED TRAINING STACK, HUB LIBRARY PLUGS REDOS HOLE

Hugging Face landed distributed training orchestration in transformers while plugging a regex vulnerability that could stall untrusted card parsing for minutes.

+10

python 61 shipped 1-min read

@huggingface 2 days ago

TRANSFORMERS HARDENS AGAINST PYTORCH FRAGMENTATION WHILE TRL SIMPLIFIES DISTILLATION

Transformers plugged a cascading import failure that breaks downstream CI on older PyTorch versions, while TRL rips out dead code to lock DistillationTrainer into prompt-only datasets.

python 70 shipped 2-min read

@huggingface 3 days ago

DATASET VIEWER LOCKS DOWN ARROW, FUNES SHIPS GROUNDED ASK

Hugging Face security teams moved overnight to contain a critical Arrow IPC parsing vulnerability in dataset-viewer while shipping three production hardening releases across Repo2RLEnv, optimum-executorch, and funes.

python 91 shipped 2-min read

@huggingface 4 days ago

TRANSFORMERS FIXES RECURRENT MASKING BUG; MOON-IDE KILLS PAIRING WINDOW

A critical padding mask bug in transformers is now fixed for chunked prefill and cache continuation, while moon-ide removes dead startup ceremony and hardens the remote bridge client.

python 97 shipped 2-min read

Elsewhere on the wire

AI Agents about 5 hours ago

CLAUDE OPUS 5 LANDS ACROSS THE STACK

The newest Anthropic model is now live in langchain, Cline, and llama-index, with native support for extended reasoning and 1M context windows.

ai-agents 28 shipped 1-min read

Local LLMs about 5 hours ago

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.