Mistral Large 4 Open Weights: 1T Model for Cyber Defense

French AI developer Mistral AI has published the weights for Mistral Large 4, a roughly 1-trillion-parameter model nicknamed “Le Chonk”, making it one of the largest openly downloadable AI systems available and the most capable to come out of Europe. The company released the weights on October 29, 2026, and is positioning the model for software engineering and, in particular, defensive cybersecurity work.

Large 4 first appeared in public preview through Mistral’s API on October 6, with the company promising open weights by the end of the month. It met that schedule.

The model

According to Mistral and its Hugging Face release:

• Architecture: a sparse mixture-of-experts design with about 1.05 trillion total parameters.
• Active parameters: roughly 49 billion per token, or about 52 billion including embeddings, depending on how they are counted.
• Context window: up to 1 million tokens, according to multiple reports.
• Formats: weights published in FP8 and NVFP4 precision.
• Hardware: the model can run on as few as four Nvidia B200 or B300 GPUs, according to coverage of the release.

Because only a small fraction of the parameters are active for any given token, inference costs are closer to those of a mid-size dense model than to a trillion-parameter one, though the full weights still need to fit in memory.

How it fits Mistral’s lineup

Large 4 sits at the top of a range that has historically split into two tracks: smaller models released with open weights, and larger flagship models offered mainly through Mistral’s API and its Le Chat assistant. Releasing the weights of the flagship itself marks a shift in that approach, and puts Mistral’s largest model in the same open category as the biggest Chinese releases.

The release also follows a period of heavy fundraising. Mistral closed a €3 billion Series D led by Samsung in September, and has expanded commercial partnerships with cloud providers and European governments. Running and serving a trillion-parameter model is expensive, and an open release lets customers absorb part of that cost on their own hardware.

Where to get it

The weights are available on Hugging Face. Mistral says organizations can deploy the model on their own infrastructure using open-source inference engines vLLM or SGLang. The model is also offered through Mistral’s own API, and reports say it is available through Nvidia NIM containers and Microsoft Foundry.

Reported API pricing varies by source, with figures ranging from about $0.70 to $1.36 per million input tokens. Mistral had not published a single consolidated price list in the materials reviewed for this article.

The licence

The weights ship under the Mistral Ecosystem Open-Weight License, which some outlets have abbreviated as the MEOW licence. It permits downloading, self-hosting and modification, but requires a separate commercial licence for companies that want to run the model as a hosted service for third parties, according to reports.

That makes Large 4 “open weight” rather than open source in the strict sense. The distinction matters mainly for cloud providers and API resellers; enterprises running the model internally are the primary audience Mistral is targeting.

The cybersecurity pitch

Mistral is leaning hard on security. The company reports a score of 51% on the Artificial Analysis Cyber Index, a benchmark that measures a model’s ability to find vulnerabilities in a codebase, reproduce and validate them, and write patches. Artificial Analysis says the index covers defensive tasks only and explicitly excludes building working exploits.

Independent testing of the preview version by Artificial Analysis scored it at 50%, according to Tom’s Hardware, which reported that the result ranked Large 4 as the best model built outside the United States and China.

Other security figures reported around the launch:

• Refusals: Mistral says Large 4 refuses 96.6% of tested malicious cybersecurity requests. This is a vendor-reported figure.
• Prompt-injection resistance: one outlet reported a 93.3% attack-resistance score on Lakera’s public B3 AI security benchmark.
• Patching: a secondary source reported an 82% success rate on a task that involves reproducing and patching real open-source vulnerabilities. That figure has not been independently confirmed.

Mistral has framed the release around sovereignty: European governments, banks and critical-infrastructure operators that want frontier-level security tooling without sending code or logs to a US or Chinese provider can now run the model on hardware they control.

Coding benchmarks

On software engineering, Mistral’s launch materials report:

• 61.7% on DeepSWE v1.1
• 59.4% on SWE-Atlas-QnA
• 28.3% on Terminal-Bench 4
• 49.8% on Artificial Analysis’s coding-agent index

Those are strong numbers for an open model, but they trail the leaders. Published comparison tables put Moonshot’s Kimi K3 at 68.0 and DeepSeek V4.1 Flash at 74.2 on DeepSWE v1.1, and VentureBeat reported that Claude Opus 5 and OpenAI’s GPT-6 Astra score about 74 on the same test, according to Startup Fortune.

Mistral itself has described Large 4 as comparable to GLM-5.3, the open-weight model from China’s Zhipu AI. Tom’s Hardware summarized the picture as Large 4 being Europe’s best effort while Chinese open-weight models remain ahead.

Dual-use concerns

The release has drawn mixed reactions from the security community. Enterprise defenders welcomed access to a capable model they can run on-premises, according to coverage, while some researchers raised concerns about dual-use risk. Unlike an API-only model, open weights cannot be recalled or patched after distribution, and safety training such as refusal behavior can be weakened through fine-tuning by anyone with the weights.

The debate mirrors choices made by US labs. OpenAI has kept its strongest cyber-capable models behind access controls and staged releases, and Chinese labs including Zhipu and Moonshot have published open weights for models with comparable or stronger coding scores.

What it means

Europe finally has a frontier-scale open model. Mistral has spent two years positioned as the European alternative to US labs, but its flagship models had lagged both American closed systems and Chinese open ones. A trillion-parameter, self-hostable model with competitive security scores gives European governments and regulated industries a credible domestic option for the first time.

Security is the wedge, not raw capability. On coding benchmarks Large 4 sits behind Kimi K3, DeepSeek and the top US models. Mistral is instead selling a combination that few competitors offer together: strong defensive cyber performance, open weights, and a European legal home. For a bank or defense ministry that cannot send source code to an external API, that bundle may matter more than a few points on DeepSWE.

Winners: European enterprises and public-sector buyers, Nvidia (whose Blackwell GPUs and NIM platform are the reference deployment path), and Microsoft, which has deepened its partnership with Mistral around sovereign AI. Security teams gain a capable model they can run against sensitive codebases without data leaving their network.

Losers: smaller security-AI vendors whose products wrap closed APIs, and US labs arguing that frontier cyber capability should remain behind gated access. Each new open release at this level weakens the case that restricting distribution meaningfully limits availability.

What to watch: independent evaluations of the final open weights against Mistral’s own figures, how quickly fine-tuned variants appear and whether any strip the refusal behavior, uptake by European governments, and whether Mistral’s licence terms hold up as cloud providers seek to host the model.

For context, see our coverage of Mistral’s $3 billion Series D, Microsoft and Mistral’s sovereign AI push, GLM-5.3’s coding and cyber benchmarks, and Moonshot’s Kimi K3.