Mistral Large 4, nicknamed "Le Chonk," is now available as open weights. Mistral opened the API preview on October 6, 2026, and the downloadable weights followed on October 29, according to Lindy's overview. Mistral's release page confirms the weights are public.

Video: Announcement video accompanying Mistral’s post about the Large 4 model and reported specifications

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

In its launch post on X, Mistral described the model as natively multimodal, with 1 trillion total parameters and 49 billion active. It called Large 4 "the best open weights model from US or Europe on aggregated benchmarks" and said open weights would follow at the end of October.

What changed between the preview and the open release

The October launch was not the finished model. SiliconANGLE says the open model went through a further three weeks of training that included customer trials. The Next Web reports that Terminal-Bench 4.0 scores rose from 28% to 41% over that stretch. ExplainX points out that the score improvements come from Mistral's own reporting.

The claims also narrowed. The October post framed Large 4 around aggregated benchmarks. Mistral's release page now says the model "ranks as the best open-weight model for coding," and adds that it is among the strongest open models overall. Moving from a broad benchmark claim to a coding claim is the most useful thing to notice when reading either announcement.

Where to download Mistral Large 4 weights and how to run them

Lindy's overview places the repository at mistralai/Mistral-Large-4-1T-A52B on Hugging Face, with a compact NVFP4 build beside the main FP8 files. The FP8 download is about 1.06 TB, spread over 54 files, according to Lindy.

Mistral's release page also names vLLM and SGLang as tools for self-hosting. A compressed variant needs four Nvidia B200 or B300 accelerators, The Next Web reports. That puts self-hosting out of reach for most individuals and small teams, so the open weights matter most to organisations with data-centre hardware.

Is Mistral Large 4 open source?

Not strictly. RankFlow puts the release in the open-weight category rather than open source, and argues the terms include a commercial condition. Mistral says its own Ecosystem Open-Weight licence governs use, and frames that licence as an adapted MIT. DevDrift warns that SaaS vendors face a particular licence condition. None of the summaries reviewed here sets out the full conditions, so the exact wording matters. Software companies building on the model should read the licence text directly.

What is still unproven about Mistral's claims

The post also says the model leads on cyber defence, manufacturing and finance workloads, and beats closed frontier models at visual grounding. It gives no scores, test sets or named competitors for them. Lindy places Large 4 seventh on Artificial Analysis's open-weights ranking, a more modest standing than the "best open weights" framing implies.

On sovereignty, the post says the model was developed in Europe and can be deployed from Europe through Mistral Cloud, but it lists no regions or data terms. That gap matters for organisations weighing where their data goes, and it is the detail to ask Mistral about before committing.

Key takeaways

  • The preview launched on October 6 and the open weights arrived on October 29, within the end-of-October window Mistral named.
  • The release is open-weight rather than open source, and its licence adds conditions to check before commercial use.
  • Most performance claims are still Mistral's own, so test the model on your own workloads before relying on them.

Frequently asked questions

Are the Mistral Large 4 weights available now?

Yes. Lindy reports the weights appeared on Hugging Face on October 29, 2026, with a compact NVFP4 variant included.

How much does the Mistral Large 4 API cost?

Sources disagree. Lindy puts the post-release price at $0.70 per million input tokens and $2.10 per million output tokens, while Tech Insider cites $1.36 for input, so check Mistral's current pricing before budgeting.

How big is Mistral Large 4?

The post gives 1 trillion parameters and 49 billion active, while SiliconANGLE reports 52 billion active. Tech Insider says it uses a mixture-of-experts architecture.

Sources