Mistral Large 4 Open Weights: The Free 1T AI Model for Security

Short answer: On October 29, 2026, Mistral released the open weights of Mistral Large 4, nicknamed “Le Chonk”. It is a Mixture-of-Experts model with about 1 trillion parameters (roughly 52 billion active per token) and a 1 million token context window. It scored 82% on the Artificial Analysis Cyber Index, a test where the model must reproduce a real vulnerability in open-source software and then patch it. You can download it from Hugging Face, but you need roughly 8 data center GPUs to run it yourself.

Watch the video:

I care about this one for two reasons. I build Laravel and Next.js apps for clients, so a model that finds and patches real bugs is interesting. And I work from Riyadh, where many companies want AI but don’t want their data to leave the country. Here is what is actually known.

What Mistral Large 4 is

Item	What’s been reported
Release of weights	October 29, 2026 (a preview came earlier in October)
Nickname	“Le Chonk”
Architecture	Mixture-of-Experts, ~1.05T total parameters, ~52B active per token (one source says 49B)
Context window	1 million tokens
Weight formats	FP8 and NVFP4 on Hugging Face
Serving	vLLM and SGLang
Also available on	Mistral’s API, OpenRouter, Microsoft Foundry, NVIDIA endpoints

Mistral also calls it the best open-weight model for coding, with about 41% on Terminal-Bench 4.0 (up from 28.3% in the October preview). That figure is from Mistral and I haven’t seen an independent test of it yet.

The security part

The headline number is the Artificial Analysis Cyber Index: 82%. This benchmark gives the model a real, known vulnerability in open-source software. To score, the model has to:

Reproduce the bug, which proves it really exists.
Write a patch that fixes it.

The Hugging Face launch blog also reports 93.3% attack resistance on Lakera’s B3 benchmark, which measures how well a model resists attacks such as prompt injection. Another write-up gives different figures (97.1% on B3 and 96.6% refusal of malicious cyber requests), so treat the exact safety numbers as vendor-reported until independent tests come out.

Why open weights cut both ways

With a closed model, the vendor decides who gets access and can switch it off. With open weights, anyone can download the model:

Defenders can run it on their own code, in their own network, with no data leaving the building.
Attackers can download the same model too.

Mistral says it red-teams the model with security firms and state authorities. One report says some vetted partners get a version with reduced moderation and expanded cyber abilities. I couldn’t confirm that on an official page, which is why the video says “reportedly”.

My take: if you run a security team, test it on your own repositories first. A model that can reproduce bugs is very good at writing proof-of-concept code, so keep it in an isolated environment and log what it does.

Can you run it yourself?

Not on a laptop, and not in Ollama on a gaming PC. Even though only ~52B parameters are active per token, all ~1.05T parameters have to sit in GPU memory:

Precision	Weights only (approx.)	Example hardware
4-bit	~525 GB	8x H100 80 GB, or 4x B200
FP8	~1.05 TB	8x H200 141 GB (tight)
BF16	~2.1 TB	multi-node

Long context adds KV-cache memory on top, so plan for headroom.

The easier route is the API. List price is about $1.36 per 1M input tokens, $0.14 per 1M cached input tokens and $4.18 per 1M output tokens. A launch sale at half price was reported to end on October 22, so check Mistral’s pricing page before you budget.

The license catch

The weights ship under the Mistral Ecosystem Open-Weight License, described as an MIT derivative. In short:

✅ Use, modify, fine-tune, and build commercial products with it.
❌ Running it as a model-as-a-service (selling hosted access to the model) needs a separate commercial license from Mistral.

If you’re an agency building an AI feature into a client’s app, you’re probably fine. If you plan to resell an API endpoint for the model, talk to Mistral first.

Saudi Arabia, the Gulf and Arabic support

Availability: the weights are a public download on Hugging Face, so a company in Saudi Arabia or anywhere in MENA can host the model on its own servers or in a local data center. That keeps data in-country, which matters for many regulated companies. I couldn’t find an official statement about API availability in Saudi Arabia specifically.
Arabic: I couldn’t find official Arabic support listed for Large 4 yet. Earlier, Mistral Large 2 listed Arabic among its strongest languages, and Mistral Saba (24B, February 2025) is Mistral’s Arabic-focused model. If Arabic quality matters for you, test Large 4 on your own Arabic prompts before you commit.

Quick checklist before you adopt it

[ ] Decide: API (easy) or self-host (data control, big hardware).
[ ] Read the license if you plan to offer the model as a service.
[ ] Test it on your own code in an isolated environment.
[ ] Test Arabic prompts if your users speak Arabic.
[ ] Re-check prices on Mistral’s pricing page.

FAQ

Is Mistral Large 4 really free?
The weights are free to download and use, including commercially. Hosting it as a paid model-as-a-service needs a separate license. The API is paid.

Can I run Mistral Large 4 on my PC?
No. It needs about 525 GB of GPU memory at 4-bit and about 1 TB at FP8, so roughly 8 data center GPUs.

What does 82% on the Cyber Index mean?
In 82% of the test tasks, the model reproduced a real vulnerability in open-source software and then patched it.

Does Mistral Large 4 support Arabic?
Official Arabic support for Large 4 isn’t confirmed yet. Mistral Saba is Mistral’s Arabic-focused model.

Is it available in Saudi Arabia?
The weights can be downloaded and self-hosted anywhere, including Saudi Arabia. I couldn’t find an official statement about API availability in the Kingdom.

Watch the video:

Sources

Mistral announcement: https://mistral.ai/news/mistral-large-4-oss/
Hugging Face blog, “Welcome Le Chonk”: https://huggingface.co/blog/mistral-large4
SiliconANGLE (Oct 29, 2026): https://siliconangle.com/2026/10/29/mistral-releases-open-weights-for-mistral-large-4-providing-sovereign-control-for-the-companys-le-chonk-ai-model/
TechCrunch (Oct 29, 2026): https://techcrunch.com/2026/10/29/mistral-releases-weights-for-its-trillion-parameter-ai-model-targeting-cybersecurity/
The Next Web: https://thenextweb.com/news/mistral-large-4-open-weights-release
Unite.ai: https://www.unite.ai/mistrals-1-trillion-parameter-le-chonk-checkpoint-lands-on-hugging-face/
DEV Community (specs, safety figures): https://dev.to/zhang_tengfei_c85cbcb9e56/mistral-large-4-specs-benchmarks-pricing-and-open-weights-102g
Hardware math: https://www.yottalabs.ai/post/mistral-large-4-hardware-requirements-gpu-memory-2026
Pricing: https://pricepertoken.com/pricing-page/model/mistral-ai-mistral-large-4
Mistral Saba: https://mistral.ai/fr/news/mistral-saba