Mistral Releases Large 4 Weights: A Trillion-Parameter Open Model

October 30, 2026 · Cortex Solutions

On October 29 France's Mistral released the Large 4 weights under a modified MIT license. Here is what it offers, where it trails, and what it means for Saudi firms.

French AI company Mistral made the weights of its flagship model, Mistral Large 4, nicknamed «Le Chonk», available for download on Hugging Face on Thursday, October 29, 2026, about three weeks after a public preview. That is according to the company's announcement and same-day coverage by TechCrunch, The Next Web and The Decoder. The model has roughly one trillion parameters, and the release is more than a technical milestone: it gives companies another option for running a large model on their own infrastructure, with conditions and real costs worth understanding first.

What Mistral announced

In its announcement, Mistral describes the model as its natively multimodal flagship, built as a mixture-of-experts system that combines instruction and reasoning. The company says it is fluent in more than 160 languages. It offers three reasoning levels: low for simple tasks, medium for everyday tasks, and high, the default, for the hardest cases that require precision.

According to Mistral's announcement, the model is available through several channels:

• Mistral's API, Mistral Studio and OpenRouter.
• Open weights on Hugging Face for deployment on your own infrastructure with vLLM or SGLang.
• Microsoft Foundry and NVIDIA-hosted endpoints.

Mistral lists API pricing at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

The license: open, with one condition

The model is released under the Mistral Ecosystem Open-Weight license, which the company describes as a modified MIT license. Mistral says the model is free to use, modify, fine-tune and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before commercial use.

The practical question for any company building a product on the model is therefore whether that product counts as such a service. It is wise to review the license text with a legal adviser before launching a commercial offering.

What the numbers say

Mistral published a set of benchmark results. These are the most notable:

• Terminal-Bench 4.0: 41%, up from 28% at preview. The Next Web reports that Mistral's chart puts GLM-5.3 at 40%.
• FrontierCode 2.0: 19.7%, the highest among the open models Mistral compared. The Decoder notes that Claude Opus 5.5 scores 46% on the same benchmark.
• Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average, according to Mistral.
• B3 Agent Security benchmark: it resisted 97.1% of attacks, against 93.3% in the preview.

The outside picture is more modest. The Decoder, citing Artificial Analysis, reports that the model scored 26 points on Intelligence Index v5, compared with 32 for both DeepSeek V4.1 Flash and GLM-5.3, and 30 for MiMo-V2.6-Pro. The publication concludes that it still falls short of the strongest open models from China. It also reports that cost per index task fell to $2.06 from $5.54 at preview. And it adds that Mistral's comparisons leave out Kimi K3.1 Pro, which scores 46 points, because its weights have not been released yet; Moonshot plans to release them by November 4.

Cybersecurity at the center of the pitch

TechCrunch presented the release as targeting cybersecurity. Pierre Stock, Mistral's VP of Science, told the outlet the model offers «access that can't be cut off by definition because the model is open-weighted». He told The Next Web that AI is going to be like electricity, and that no company should depend on someone who could cut off its access overnight.

Mistral says the model refused 96.6% of malicious cyber requests on average, the highest it measured among open models. On the Artificial Analysis Cyber Index it scored 51%. However, The Decoder points out that Mistral's own chart puts DeepSeek V4.1 Flash at 57% and MiMo-V2.6-Pro at 56%, which makes Large 4 third among open models rather than first.

Mistral also has a specialized variant for penetration testing and vulnerability assessment, available in preview to design partners such as Orange Cyberdefense, subject to vetting and contractual safeguards. The company acknowledges it sees a potential for uplift to conduct real attacks, and says it is modifying its guardrails accordingly.

Self-hosting is not simple

Open weights do not mean easy deployment. The Decoder, citing the SGLang documentation, reports that the standard version runs on eight Nvidia B300 GPUs, while a more heavily compressed variant runs on four GB300 GPUs. The Next Web reports a quantised version running on four B200 or B300 GPUs. Mistral says a full technical report is coming soon, and The Next Web reports that smaller distilled versions are planned, including one that runs on a laptop, without committed dates.

What this means for Saudi businesses

For Saudi companies, the announcement does not mean rushing to adopt a new model. It does widen the options, and these are the practical takeaways:

• Self-hosting becomes more relevant for sectors that handle sensitive data, because it keeps data within infrastructure the company controls. It does not by itself guarantee compliance with the Personal Data Protection Law or sector rules; that depends on how the system is designed and operated.
• The hardware requirements put self-hosting out of reach for most small and medium businesses at this stage, so the API or a smaller model is the more realistic route.
• Compare cost per completed task, not just price per token. The figures above show that cheaper tokens and a cheaper task are not the same thing.
• Test on your own work in Arabic. Support for more than 160 languages says nothing about quality on your contracts, customer messages or dialect.
• Treat vendor benchmarks as a starting point, and wait for independent evaluations and the technical report before making a large decision.
• Design your applications so a model can be swapped easily. The gap between models is changing within weeks, and Kimi K3.1 Pro's weights alone may reshuffle the open-model ranking next week.