Mistral Large 4 — “le Chonk” in Mistral’s own words — is the French lab’s October 6, 2026 public preview of its biggest model yet: a 1.05-trillion-parameter multimodal mixture-of-experts with 52 billion active parameters and a 1M-token context window. The weights are promised for the end of the month, and it is the first milestone on the roadmap funded by Mistral’s €3 billion Series D from September 8.
The pitch is not raw leaderboard rank. Mistral pairs “competitive with the strongest open-source models globally” with a sovereignty story: trained on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers, served from the same infrastructure, and deployable in a European region Mistral operates under European law.
Benchmarks
All figures are Mistral-reported (October 6, 2026); weights are not out, so nothing here has been independently reproduced yet.
| Area | Result |
|---|---|
| DeepSWE v1.1 (coding) | 61.7% |
| SWE-Atlas-QnA (repository understanding) | 59.4% |
| Terminal-Bench 4.0 | 28.3% |
| Coding Agent Index | 49.8% — ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max |
| AutomationBench (657 business workflows) | 59.9% — ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro |
| AA-Briefcase (long-horizon knowledge work) | 1,393 Elo — ahead of DeepSeek V4 Pro |
| Artificial Analysis Cyber Index | Top five of all models; leads non-Chinese open-weight models by a wide margin |
| Vulnerability reproduce-and-patch task | 82% — highest of any model, per Mistral |
| Cybench | 93% of 40 challenges |
| Dense 200 (visual grounding) | 42% vs GPT-6 Astra 41% |
| Lakera B3 attack resistance | 93.3% |
Two other results are worth knowing. In a blind Surge AI human evaluation of coding quality, Mistral Large 4 Preview ranked second of five (3.74), behind only Claude Opus 5 (4.22) and ahead of GLM-5.3 (3.60), Kimi K3 (3.59) and GLM-5.2 (3.40). In internal expert comparisons against GLM-5.3 it was preferred in CAD and STEM and roughly on par in finance and coding. Third-party evaluators at Vals also found it ahead of GPT-6 Astra on its finance and legal agent tasks, per Mistral.
Pricing
| Per 1M tokens | List price | Preview sale (50% off) |
|---|---|---|
| Input | $1.36 | $0.68 |
| Cached input (as listed) | $0.14 | $0.07 |
| Output | $4.18 | $2.09 |
That undercuts Grok 4.7, Claude Sonnet 5.5 and GPT-6.1 Sol on list price, but not DeepSeek V4.1 Flash or the Qwen3.8 Flash models. Preview pricing is provisional.
Specs
| Spec | Value |
|---|---|
| API model | mistral-large-4 (version 26.10) |
| Parameters | 1.05T total, 52B active, 1.6B vision encoder |
| Context window | 1M tokens |
| Modalities | Text and image in, text out |
| Reasoning | Hybrid instruct-and-reasoning in a single model |
| Languages | 160+ in training data, including all official EU languages |
| Features | Function calling, structured outputs, document QnA, agents and conversations, built-in tools, batch API |
| Weights | Promised by the end of October 2026 |
Safety
Mistral reports that the model saturates its indirect-prompt-injection benchmarks, scores 1.691 of 2 on the KORA responsibility benchmark (its highest among open models), and refuses malicious cyber prompts (JailbreakBench, StrongREJECT, AgentHarm) more often than any open model it compared, while still scoring at the top on legitimate cyber tasks. Vetted partners and state authorities are red-teaming a version with reduced moderation and expanded cyber capabilities before release.
How it compares
- vs DeepSeek V4.1 Flash: Mistral’s DeepSWE (61.7%) sits well below DeepSeek’s reported 74.2% (different harnesses), and V4.1 Flash is much cheaper. Mistral’s edge is cyber, legal and sovereignty positioning.
- vs GLM-5.3: Mistral’s own human eval prefers Large 4 in CAD and STEM and calls finance and coding close; GLM-5.3 posts higher public coding numbers.
- vs Claude Opus 5.5 and GPT-6: closed models lead on coding and agentic benchmarks; Mistral’s claim is parity on enterprise verticals and better access for security work.
Limitations
- Preview only. No weights, license or full technical report yet.
- Coding gap. Its coding scores sit well below the leading open and closed models.
- Vendor-reported. Cyber comparisons depend on closed models refusing the test task.
See the 2026 LLM guide for where open-weight models stand overall.
Best for
- Sovereign and on-premises deployments in regulated industries (finance, public sector, defense)
- Cybersecurity work — vulnerability reproduction, patching, malware analysis and detection engineering
- Document, chart and technical-drawing understanding, including visual grounding
- Multilingual enterprise assistants across the EU and 160+ languages
Pros & cons
Strengths
- A 1.05T-parameter MoE with 52B active parameters, native image understanding and a 1M-token context — Mistral's largest model and, by its own claim, the strongest open-weight model developed in the US or Europe
- Leads on enterprise-sensitive work — Mistral reports state-of-the-art among open models on cybersecurity, finance and law, and a top-five spot on the Artificial Analysis Cyber Index
- Open weights promised for the end of October, with self-hosting and a European deployment Mistral runs under European law
- Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, with 160+ languages in the training data including every official EU language
Limitations
- It is a public preview — weights, a license and the full architecture and post-training details are still to come
- Coding scores trail the Chinese open-weight leaders and every closed frontier model Mistral compares against (DeepSWE v1.1 61.7%, Terminal-Bench 4.0 28.3%)
- Almost all numbers are Mistral-reported or taken from a small human evaluation, and the cyber comparison leans on closed models that refuse the test task
- A 1T-parameter MoE will need a multi-GPU node to self-host
Frequently asked questions
When was Mistral Large 4 released?
Mistral launched a public preview on October 6, 2026, with the API available on Mistral Studio. The weights are due by the end of October 2026, so as of mid-October it is not yet downloadable.
How big is Mistral Large 4?
Mistral's docs list 1.05 trillion total parameters, 52 billion active per token and a 1.6-billion-parameter vision encoder. It is a granular mixture-of-experts model with native multimodal input and a 1M-token context window.
How much does Mistral Large 4 cost?
Mistral's model page lists $1.36 per 1M input tokens and $4.18 per 1M output tokens, with a 50% preview sale currently bringing that to $0.68 and $2.09. Treat preview prices as provisional.
Is Mistral Large 4 really open source?
Mistral calls it open-weight and says it will release the weights by the end of October. A license has not been published, and the preview itself runs on Mistral's hosted API, so check the license terms when the weights drop before building on it.
Why does Mistral say closed models score near zero on its cyber test?
On one Artificial Analysis Cyber Index task — reproduce a real vulnerability and then patch it — Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task, while Mistral Large 4 scores 82%. That is a statement about refusals as much as capability, and it comes from Mistral's own write-up.