Mistral’s Large 4 ‘Le Chonk’ joins the West’s open model push

Mistral introduced Large 4 on October 6, opening a public API preview of the model it calls “Le Chonk.” The model has roughly one trillion parameters, and Mistral says it outperforms competing open models from U.S. and European labs.

It follows Reflection’s October 5 announcement of Beam and extends the Western open model push covered in The Rundown. Mistral plans to publish downloadable model weights by the end of October. Reflection is taking registrations for early access and promises its weights under Apache 2.0 later this month.

What the tests show

Mistral reports 61.7% on DeepSWE v1.1, a coding benchmark. Reflection’s published comparison puts Beam at 44.4%, Kimi K3 at 68.0% and DeepSeek V4.1 Flash at 74.2%. Those figures place Large 4 above Beam and below the two Chinese models, though the reports may reflect different test conditions.

An evaluation from Artificial Analysis gives Large 4 an 82% score on CyberGym-E2E-AA, ahead of the models it compared. That offers outside support for Mistral’s claim that cybersecurity is a strength.

Legal testing also looks promising. Vals’ October 6 results table lists a 15.83% task pass rate for Large 4 on Harvey’s Legal Agent Benchmark, against 5.42% for GPT-6 Astra. That is nearly three times the score on this specific test, with a low absolute completion rate that deserves attention.

Why it matters

Le Chonk deserves the prize for AI’s best name. Alongside Beam, it also gives the West’s open model push a reason for optimism. Chinese labs still lead important open model comparisons, but two serious Western announcements in one week make the market look more competitive.

For organizations handling sensitive code or legal documents, Mistral could become another option for keeping work on infrastructure they choose. Mistral describes infrastructure it operates in Europe and plans for private cloud deployments and installations on customers’ premises. Teams can test the API preview now. Running Large 4 themselves will depend on the public weight release, its license, software support and hardware requirements.

For legal teams, the relative gain offers a reason to try supervised document work. Vals requires every criterion to be met for a task to pass and averages scores from two AI judges. Large 4’s low absolute pass rate calls for close human review of its output. A useful specialist model can earn consideration even while other models lead broader comparisons.

For security teams, usefulness also depends on which tasks a model will perform. Mistral says selected cyber partners and state authorities receive reduced moderation and expanded cyber capabilities during the preview. It attributes some rivals’ poor results on the cited cyber test to refusals. The broader permissions could help legitimate defenders. How Mistral grants and audits that access remains an important safety question.

The new options also have to compete on cost. Artificial Analysis measured Large 4 at $1.13 per Intelligence Index task, or $0.57 with the launch discount. GLM-5.3-Flash cost $0.25 and DeepSeek V4.1 Flash $0.27 on that measure. Those figures cover the evaluator’s workload. Buyers should test cost and reliability on their own work, along with the control each deployment option gives them.