Mistral Large 4 “Le Chonk” Claims the Open-Weight Lead in Coding and Cyber Defense

France’s Mistral says its Mistral Large 4, nicknamed Le Chonk, now matches the best Chinese open models at agentic coding and cybersecurity. The weights are free to download, so companies and governments can run the model themselves. Short version: Europe has a contender again, and early independent tests from Artificial Analysis and Vals.ai now let us check the company claims.

What Mistral Actually Released

Mistral first previewed Mistral Large 4 on October 6, 2026. On October 29, it published the open weights of the finished version. “Open weights” means the trained numerical parameters are public. Anyone can download them, host the model on their own servers, and fine-tune it. That is different from open source software in the strict sense, but for users the practical effect is control: no API shutoff, no data leaving your network.

The model is large. Mistral reports 1 trillion total parameters with about 49 billion active per token (52 billion counting embeddings), a mixture of experts design where only part of the network runs on each step. Inference, the process of generating answers from a trained model, is therefore cheaper than the headline size suggests. The model is natively multimodal with a claimed 1M token context window and offers three reasoning effort settings: low, medium, and high. It ships under the Mistral Ecosystem Open-Weight License, which allows free use and fine-tuning but requires a separate commercial license for model-as-a-service offerings.

The Benchmark Claims, With Context

Mistral says three extra weeks of reinforcement training on harder coding and security tasks lifted the model well past its October 6 preview. Benchmarks are standardized tests used to compare models. They are useful but narrow: they reward the exact tasks they measure, not general reliability. Treat Mistral’s own figures as its best case until more independent scores land.

The pattern in the numbers is consistent. Mistral claims the lead on the two coding tests it highlights, ties the composite coding index, and trails on overall cyber and intelligence measures. Both headline coding figures come in lower in independent runs: Terminal-Bench 4.0 at 25 to 26 percent versus the claimed 41, and FrontierCode 2.0 at 15 percent versus the claimed 19.7. Scores move between testing setups, and no single number settles the race.
Mistral’s vendor numbers also include figures the independent testers have not restated. On FrontierCode 2.0, the company claims 19.7 percent against 19.3 for DeepSeek V4.1 Flash and 17.8 for GLM-5.3, a narrow claimed lead. On the composite Coding Agent Index it claims 48 against GLM-5.3’s 49 and DeepSeek’s 46, which reads as a tie at the top. On efficiency, Mistral claims Le Chonk needs about 47 percent fewer tokens per task than GLM-5.3 for comparable results. Its claimed 82 percent on the reproduce-then-patch cyber test, where closed frontier models often refuse outright, also remains unverified.

Independent Vals.ai Scores: Open-Model Comparison

Mistral-only Vals results: Legal Index 67.68%, Finance Index 56.18%, Tax Agent Bench 27.33%, HLAB-Vals-1.1 15.83%. Vals.ai has not tested Muse Spark Open.

Independent Artificial Analysis Scores: Open-Model Comparison

The three rival columns were measured under Artificial Analysis v4.3.2, while Mistral Large 4 scores under v5.0, so the Intelligence Index cells for the rivals are left blank rather than compared across index versions. Per-test scores are comparable, with the CritPt version difference noted in-row. Mistral-only AA results: Strategy & Ops Index 43, Finance & Accounting Index 39, Openness Index 39, Economics Index 37, Legal Index 35, Engineering Index 33, AA-Omniscience accuracy 26 percent with a 72 percent non-hallucination rate, HLE-Diamond 17 percent, and Terminal-Bench-Science 0.1 at 4 percent.

On the shared Artificial Analysis tests, DeepSeek V4.1 Flash still leads Mistral in four of five, which is the clearest independent evidence that the open-weight lead has not changed hands. The independent picture is sharper than the launch claims. Vibe Code Bench at 81.56 percent is the standout, and legal and finance agent scores are solid. But Terminal-Bench lands at 25 to 26 percent in both independent runs against the claimed 41, and FrontierCode 2.0 comes in at 15 percent independently versus the claimed 19.7. The two headline agentic coding numbers are the ones third party testers could not reproduce. Cyber Index overall sits at 51 on Artificial Analysis, behind DeepSeek V4.1 Flash at 57, and the claimed 82 percent reproduce then patch score remains unverified. Hallucination behavior looks reasonable: 72 percent non-hallucination on AA-Omniscience. Artificial Analysis also clocks the model at 161 output tokens per second, among the faster open releases at this size.

On cost, Mistral claims Le Chonk uses about 47 percent fewer tokens per task than GLM-5.3 for comparable performance. Fewer tokens means lower bills and faster runs. That efficiency claim, if confirmed by third parties, could matter more to everyday users than a one-point benchmark edge.

Can You Run Mistral Large 4 Locally?

Not on a laptop. A trillion-parameter model needs data center-grade GPUs even in compressed form, and the official deployment recipes target 8 GPU servers. What open weights do allow is self-hosting on rented or on-premises clusters, plus fine-tuning for specific codebases. For background on how smaller open models serve local use, see our coverage of MiniCPM5-2B and the under-4B intelligence index.

How It Compares With Chinese Open-Weight Rivals

Until this release, Chinese labs set the pace in open weights. Models from DeepSeek and Z.ai ranked at or near the top of public coding and reasoning tables through most of 2026. For context on that ecosystem, read how China’s open-weight AI ecosystem advantage compares with US subscription APIs.

Mistral’s framing is direct: Le Chonk matches DeepSeek and Z.ai flagships on coding and cyber tasks while costing less to run. One correction to the leapfrog narrative: on the composite Coding Agent Index, it ties rather than leads GLM-5.3, and closed models from Anthropic and OpenAI remain well ahead overall. Chinese rivals still hold depth across model sizes, from small efficient releases such as Qwen 3.8 at 27B to frontier scale systems. One release does not reset the whole field. It narrows the gap at the top for two specific workloads.

Why Cybersecurity Is the Center of This Story

Mistral pitches Le Chonk as built for defenders: finding bugs, triaging alerts, and writing patches inside networks that cannot send code to a foreign API. Recent incidents give that pitch urgency. As WIRED reported on October 29, an OpenAI model escaped a test sandbox over the summer, and AI-assisted vulnerability hunting knocked sites offline in October after flaws in web infrastructure came to light.

Open weights cut both ways. Defenders get a model they can audit and keep offline. Attackers can download the same weights.

Sophos threat intelligence lead Rafe Pilling told WIRED it is clearly dual-use technology and that monitoring and access controls matter regardless of where a model was trained. That caution is worth keeping in mind alongside Mistral’s sovereignty argument that open weights cannot be switched off remotely.

That sovereignty point also explains the European angle. Training ran on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centers, according to the company. For governments weighing dependence on US closed APIs or Chinese weights, a capable European option changes procurement conversations even before every benchmark is verified.

What Changed and What Did Not

What changed: a Western lab claims the open model lead on agentic coding, with detailed efficiency numbers and downloadable weights, and early independent tests back the everyday coding story (Vibe Code Bench 81.56 percent) while undercutting both headline agentic figures. What did not: broader reasoning and multilingual comparisons, and the day-to-day reliability questions that benchmarks miss.

Mistral’s technical notes are the place to check primary details, starting with Mistral’s official newsroom. If the claimed efficiency holds up under independent testing, Le Chonk gives European teams a practical alternative. If it does not, the race narrative resets again. Either way, users gain from weights they can inspect and host.