Hugging Face AI Updates: October 31, 2026

1. Transformers v5.22 Adds Mistral Large 4 and Nemotron 3.5 Transcribe, Enables LoRA Training on FP8 Checkpoints

Hugging Face released transformers v5.22.0, headlined by Mistral Large 4 (“Le Chonk”), a roughly 1T-parameter granular MoE with 128 routed experts (4 active per token) plus one shared expert, a Pixtral-style vision encoder, a 1M-token context and per-request reasoning effort set to low, medium or high (the default). The notes recommend temperature 1.0 and top_p 0.95 and give the active parameter count as 49B per token, or about 52B including embeddings and output layers. The release also adds NVIDIA’s Nemotron 3.5 Transcribe speech model and left-context support for Nemotron3Diarization. The training-side change is that FP8-quantized models can now be used for adapter training with PEFT (is_trainable is now True), with backward passes added for dense and grouped FP8 matmuls; experts are dequantized in chunks and autograd is skipped at inference, cutting peak memory in the MoE layer backward from about 28.9 GiB to 2.8 GiB at 8k tokens. Expert-parallel logic moved into its own file, which the notes list as a breaking change. Pipelines now defer eager model imports and exporters resolve patch paths lazily to reduce import overhead. The v5.22.1 patch on October 30 only removes a work-in-progress documentation page. Source Source

2. TRL v1.17 Fine-Tunes Mistral Large 4 on One 8xB300 Node, Fixes vLLM Importance Sampling Bias and Makes Score Centering the AsyncGRPO Default

Hugging Face released TRL v1.17.0. It supports LoRA SFT of Mistral Large 4 (1T total, 52B active) from its FP8 weights on a single 8xB300 node, at about 32.7k tokens per step in roughly 8 seconds using 142 GiB per GPU, and requires transformers 5.22.0, peft 0.21.0 and accelerate 1.15.0 or newer. The vLLM importance-sampling ratio is now normalized over the candidate set vLLM actually sampled from under top-k, top-p or min-p rather than the full vocabulary; the previous ratio was biased toward zero, and GRPOConfig now errors if top_p < 1.0 or min_p > 0.0 is set without top_k > 0 unless vllm_importance_sampling_correction=False. AsyncGRPO switches to score centering by default (importance_sampling_min=0.5, importance_sampling_max=5.0, tokens outside the range dropped), which costs 15 to 20% generation throughput; set score_centering=False to keep the clipped objective. AsyncGRPO also gains an adaptive in-flight cap tied to vLLM KV cache usage, filters zero-advantage groups and cuts TreeRow.pack from 250 s to 31 s on a 15-step run. Other additions: tree attention for the linear-attention hybrids Qwen3.5, Qwen3.6, Qwen3-Next and OLMo-Hybrid, training on AMD GPUs through PyTorch’s ROCm build with a guide and CI, GOLD X-Token cross-tokenizer distillation (including byte-fallback tokenizers such as Gemma), Beam model support, a Mistral Medium 3.5 chat template, LFM2.5 tool calling and vLLM 0.32.0. The experimental TPO trainer is removed, dependency floors rise to torch 2.6.0, accelerate 1.11.0 and peft 0.17.0, liger-kernel is dropped from dev dependencies, and the async trainers now save checkpoints in the mixed-precision dtype rather than fp32. The v1.17.1 patch on October 30 stops a parenthesis inside an LFM2.5 tool-call argument from closing the call early. Source Source

3. huggingface_hub v2.4.0 Copies Files Into Buckets Through HfFileSystem and Flags Outdated CLI Extensions

Hugging Face released huggingface_hub v2.4.0. HfFileSystem can now copy into a Bucket from the same Bucket, another Bucket or a model, dataset or Space repo; Xet-stored files are copied on the Hub by hash with no download, other files are downloaded and re-uploaded, and mv and rename work inside a Bucket as a result. Copying from a Bucket into a repo is not yet supported. hf extensions list gains VERSION and LATEST columns, the CLI prints a once-a-day hint when an extension is outdated (disable with HF_HUB_DISABLE_UPDATE_CHECK=1), and hf extensions update rebuilds a Python extension’s virtualenv when its interpreter no longer starts, such as after a Homebrew upgrade. create_webhook now watches both repo and discussion domains when domains is omitted. Fixes include try_to_load_from_cache rejecting filenames containing ../ or absolute paths that could resolve outside the cache folder, HfFileSystem.find(detail=True) returning info for a single file path (which had broken datatrove readers), and hf cache verify --local-dir no longer reporting .cache/huggingface/ metadata as extra files. Source