Transformers v5.22.0, released October 29, 2026, adds support for Mistral Large 4 (nicknamed "Le Chonk"), which Mistral describes as its largest and most capable model to date. Read the release notes.
What the announcement establishes
- Architecture: a 1-trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers) — 128 routed experts, 4 active per token, plus 1 shared expert.
- Context and languages: a 1M-token context window and native fluency in 160+ languages, including every official EU language.
- Modalities and modes: text and image input with text output (Pixtral-style vision encoder, variable image sizes); reasoning effort selectable per request at low, medium or high, with high as the default.
- Training: Mistral states the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters.
Who is affected and how to use it
The weights target self-hosting developers via transformers (and the surrounding inference stack). Mistral's recommended settings: temperature=1.0, top_p=0.95 for all reasoning efforts, well-scoped tool sets and clear system prompts for agentic use.
License conditions
The release uses the Mistral Ecosystem Open-Weight License, a modified MIT license permitting commercial and non-commercial use — but operating a Model-as-a-Service business on top of it requires a separate commercial license from Mistral.
What remains a vendor claim
Mistral reports strong performance on cybersecurity, finance and law workloads, state-of-the-art visual grounding among open-weight models, and cybersecurity results described as state of the art among open weights. These figures come from the vendor's release notes, not independent evaluation; no benchmark tables accompanied the transformers release.
Also in this release
v5.22.0 lists breaking changes, including separation of expert-parallelism logic into its own file, and a v5.21.1 companion patch fixed a rope-scaling setter regression.
Recent library coverage: v5.21.0 and v5.20.0. Vendor-benchmark caveats are explained in our verification guide.