Recommended Articles
Tech & AI

Mistral Large 4 Open Weights: A Trillion-Parameter Model Anyone Can Download

By Angeled · October 30, 2026 · Tech & AI
Mistral Large 4 Open Weights: A Trillion-Parameter Model Anyone Can Download
Mistral Large 4 Open Weights: A Trillion-Parameter Model Anyone Can Download

Paris-based Mistral AI released the open weights for Mistral Large 4 on October 29, 2026, making its flagship artificial intelligence model — nicknamed "Le Chonk" — downloadable and runnable on independent infrastructure. The release marks a milestone for open-weight AI: a trillion-parameter, frontier-class model that organizations can inspect, customize, and deploy on their own servers rather than accessing only through a hosted API.

What was released

Mistral Large 4 is a natively multimodal mixture-of-experts model with roughly 1 trillion total parameters, about 52 billion of them active for each token (49 billion excluding embeddings and output layers). It uses 128 routed experts with four active per token plus one shared expert, accepts text and image input, and supports a context window of 1 million tokens — the amount of material the model can consider in a single request.

The weights landed three weeks after the model's October 6 public preview. Mistral says the released version reflects three additional weeks of reinforcement learning, which lifted its Terminal-Bench 4.0 coding score from 28.3% at preview to 41%. The weights are available on Hugging Face, with deployment support across tools including vLLM, SGLang, Transformers, and llama.cpp. The model is also served through the Mistral API and Mistral Studio, OpenRouter, NVIDIA-hosted endpoints with an optimized NIM container, and Microsoft Foundry, the first cloud to carry it.

Built for coding and cyber defense

Mistral is pitching Large 4 primarily at coding and cybersecurity work. The company reports a score of 19.7% on Cognition's private FrontierCode 2.0 benchmark, the top open-weight score, and 48% on the Artificial Analysis Coding Agent Index — effectively tied with China's GLM-5.3 while using 47% fewer tokens per task. On the Artificial Analysis Intelligence Index v5, Large 4 scores 26, the highest among Western open-weight models, though still behind Chinese open models such as DeepSeek V4.1 Flash and GLM-5.3.

On cybersecurity, Mistral says the model scores 51% on the Artificial Analysis Cyber Index and refused 96.6% of malicious cyber requests in its tests. A more aggressive penetration-testing version stays restricted to vetted partners under contractual restrictions, with partners including Orange Cyberdefense testing it. Mistral acknowledges the model could make real attacks easier and says it plans to tighten safeguards in the coming months.

The license: open, with one big carve-out

The weights are published under the Mistral Ecosystem Open-Weight License, a modified MIT license that permits free use, modification, fine-tuning, and commercialization. The catch: companies that offer Large 4 as a hosted model service — a Model-as-a-Service business — need a separate commercial license from Mistral. Internal use and access through Mistral's official products are not affected.

API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — roughly half the preview price. Mistral says the model was trained from scratch on its own European data centers, using around 4,000 NVIDIA Grace Blackwell GPUs over two months, according to company statements reported by ZDNET, and is natively fluent in more than 160 languages.

What "open" still costs

Downloading the weights is free; running them is not trivial. Self-hosting guidance indicates the standard version needs eight high-end NVIDIA GPUs, with a more compressed variant running on four. That makes Large 4 accessible to well-resourced enterprises, research labs, and sovereign AI programs, but still out of reach for a laptop or a small startup's spare server — a reminder that open weights lower the software barrier far more than the hardware one.

Why it matters for founders and investors

For founders, Large 4 changes the build-versus-buy calculus for AI products. A team can now self-host a frontier-class coding and security model, fine-tune it on proprietary data, and keep sensitive workloads in-house — a pitch that resonates with regulated industries and European customers pursuing sovereign AI. But the license carve-out matters: anyone planning to resell model access as a service needs Mistral's separate permission, which limits one obvious business model. For investors, the release sharpens the competitive question facing every AI lab: if trillion-parameter models can be downloaded for free, the durable value shifts from the model itself to distribution, data, workflow integration, and trust.

Angeled.com connects founders with investors.

Mistral Large 4 open weights coverage, Unite.AI, October 29, 2026
Mistral Large 4 open weights coverage, Skimwire, October 29, 2026
Mistral Large 4 release coverage, TestingCatalog, October 29, 2026
Mistral Large 4 benchmark analysis, Pivot News, October 30, 2026

Read Next