$ cat huggingface/week/2026-10-05.log
the week in review · Oct 5 – Oct 11, 2026
Diffusers 0.41.0 ships Qwen-Image 2.1, TRL 1.15 goes breaking
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
Diffusers 0.41.0 adds Qwen-Image 2.1 pipelines for generation, editing, transparency and LoRA training, while TRL 1.15 fuses its LM head and drops Python 3.10.
v1.15.0 huggingface/trl
SFT, DPO, KTO, GRPO, RLOO and Distillation now score tokens through the fused LM head instead of the old chunked cross-entropy, which the release notes put at "up to 6.9× longer sequences on the same GPU". SFT's Liger loss path and its own chunked cross-entropy are gone, and with Python 3.10 dropped, existing training scripts need a dependency and runtime check before this upgrade.
Diffusers 0.41.0: QwenImage 2.1 pipeline and more huggingface/diffusers
Qwen-Image 2.1 arrives with text-to-image, image editing, native transparency and LoRA training, and diffusers now follows the Transformers release model: minor releases coordinated around new model integrations, with patches in between.
[hf-xet v1.7.0]: Telemetry, Dynamic Download Buffers, and fixes huggingface/xet-core
hf-xet gains client-side telemetry for every upload and download and sizes download buffers from available RAM. A shard's key and ETag now depend only on its content, since CAS no longer lets the footer timestamp give identical bytes a different S3 object per upload.
End completions on every eos id the model declares in GRPO, RLOO and Distillation huggingface/trl
GRPO, RLOO and Distillation now end a completion on any eos id the model declares in its generation config, not just the tokenizer's eos token. The PR names Phi-3.5 and Gemma 3/4 as models that end turns on ids the tokenizer does not own.
Recognize the current LFM2 chat template for assistant-only loss huggingface/trl
SFTTrainer with assistant_only_loss=True works again on LiquidAI/LFM2-1.2B and the LFM2 checkpoints sharing its template, after that template changed. Companion fixes cover the current LFM2.5 template and every Gemma 4 template revision shipped by google/gemma-4-E2B-it, so GRPOTrainer with tools works on a pinned revision.
Drop Python 3.10 support (#7497) huggingface/trl
Python 3.10 is no longer supported anywhere in the library. The build and CI matrix now assume 3.11 or newer.
Kandinsky6 (#14949) huggingface/diffusers
Kandinsky 6 lands with TI2VA and SR pipelines plus PiflowScheduler and MMAudioVAE, and the model code drops its bundled bigvgan and manual checkpoint loading.
feat: add machine readability to kernels variants huggingface/kernels
kernels variants can now emit machine-readable output, so compatibility checks over kernels-community repos can run in scripts rather than by eye. Alongside: transformers redirects deprecated pipelines and raises on removed ones, swaps RoPE's inner-dimension-1 matmul for a broadcast multiply, fixes quantized-cache config mutation and reorder_cache, huggingface.js inference-v4.13.32 wires OpenRo...
$ ls huggingface/week/ # the briefings behind this review
Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.