Mistral previews Large 4 ahead of planned weight release

Mistral launched Large 4 as a public preview under the code name “Le Chonk.”[1] The company plans a roughly three-week testing period with developers, cybersecurity leaders and government authorities.[1] After that period, Mistral plans to offer the model through its API and publish its weights on October 27.[1] The weights are expected to use a custom Mistral license.[1] VentureBeat reported that Mistral had not provided API pricing in the materials it reviewed, so teams assessing deployment costs will need to wait for more information.[1]

Large 4 has one trillion parameters in total, of which 49 billion are active.[1] Mistral says it trained the model from scratch over roughly two months using 4,000 Nvidia Grace Blackwell GPUs in its European data centers.[1] The company says it trained the model across more than 160 languages, including every official language of the European Union.[1] Mistral says the model accepts multimodal inputs but generates text output.[1] That distinction matters for teams considering image-heavy workflows: the stated input capability does not mean the model produces image outputs.[1]

Mistral reported a preliminary 62% score for Large 4 on the DeepSWE v1.1 software-engineering benchmark.[1] Its chart listed GLM-5.3 at 61%, but the live leaderboard listed GLM-5.3 and Kimi K3 at about 69%, with GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 around 74%.[1] On Finch, a finance-and-accounting workflow benchmark, Mistral reported 67%, tied with DeepSeek V4 Pro 0813 and ahead of GLM-5.3 at 65%.[1] Mistral’s chart showed a 15% task-pass rate on Harvey’s Legal Agent Benchmark.[1] The reported figures therefore point to different results across tasks, rather than a single uniform performance level.[1]

Mistral is targeting Large 4 for software engineering, cyber defense, financial analysis, satellite and aerial imagery, technical drawings and chip design.[1] The company says the model is particularly effective at cyber, coding, manufacturing, finance and multimodal tasks.[2] Mistral also says its cyber-defense capabilities could help enterprises and governments defend against threat actors using jailbroken closed models for cyberattacks.[2] Those are the company’s stated use cases and claims; teams evaluating the preview can test them against their own workflows before deciding whether the planned API or weight release fits their needs.[1]