Mistral Large 4: Specs, Benchmarks, Pricing and Open Weights

Mistral AI released Mistral Large 4 as a public preview on October 6, 2026, and put the weights on Hugging Face on October 29. It's a 1.05T-parameter mixture-of-experts model, the largest open-weight model any US or European lab has released, and on the day the weights shipped Mistral cut the API price roughly in half.

Mistral calls it ML4 for short, and its nickname is "Le Chonk".

Quick answer

• 1.05T total parameters, about 52B active per token, mixture-of-experts
• 1M-token context window, text and image input, trained on data in more than 160 languages
• Open weights on Hugging Face since October 29, in FP8 (about 1.06 TB) and NVFP4 (about 610 GB)
• API price from October 29: $0.70 input, $0.07 cached input and $2.10 output per million tokens
• Model ID: mistral-large-4-0, alias mistral-large-4
• The strongest open-weight model from outside China on the Artificial Analysis Intelligence Index, though leading Chinese open models still score higher

Specs

	Mistral Large 4
Total parameters	1.05T
Active parameters	About 52B (49B per token excluding embeddings and output layers)
Architecture	Granular mixture-of-experts: 128 routed experts, 4 active per token, plus 1 shared expert
Context window	1M tokens
Input	Text and images, through a 1.6B-parameter vision encoder
Languages	Training data in more than 160 languages, including every official EU language
Training	3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters

Mistral's announcement and its docs gave two different active-parameter counts at launch, 49B and 52B. The model card explains the gap: 49B parameters run per token, and 52B includes the embedding and output layers.

The mixture-of-experts design is how a 1T-parameter model can sell for $0.70 per million input tokens. The total parameter count sets how much the model can know, while the active count sets the compute each token costs. Only about 5% of the weights run per token, so ML4 costs roughly what a 50B model costs to serve.

The same model handles quick answers and step-by-step reasoning, so you don't need a separate reasoning model. Reasoning tokens count toward max_completion_tokens, though, and higher reasoning settings raise output costs.

Benchmarks

Mistral says ML4 "pushes the frontier of open-weight performance" and calls it the best open-weight model from the US or Europe on aggregated benchmarks. Most of the published numbers are Mistral's own.

Security is the strongest area. ML4 is in the top five on the Artificial Analysis Cyber Index, scores 82% on a test where the model has to reproduce a real vulnerability and then patch it, and solves 93% of Cybench challenges. Mistral is red-teaming it with security firms, vetted partners and state authorities, and some of them get a version with reduced moderation and expanded cyber capabilities.

On agent coding, ML4 scores 41% on Terminal-Bench 4.0, up from 28.3% in the preview. On Dense 200, a visual grounding test, it scores 42%, a point ahead of GPT-6 Astra. In Mistral's safety testing it resisted 97.1% of attacks on Lakera's B3 benchmark and refused 96.6% of malicious cyber requests. Mistral also claims top results on finance and legal benchmarks.

Where it ranks independently

Artificial Analysis said the launch makes "France home to the most intelligent model outside the US and China." On its Intelligence Index (v4.3.2, at preview), ML4 scored 38, tied with GPT-6 Luna and in 18th place among the 25 models shown. That puts it well ahead of other Western open-weight models, such as Nemotron 3 Ultra at 23, but behind the leading Chinese open models:

Open-weight model	AA Intelligence Index (v4.3.2)
MiMo-V2.6-Pro	46
GLM-5.3	45
Kimi K3	44
GLM-5.3-Flash	42
DeepSeek V4.1 Flash	39
Mistral Large 4	38
Nemotron 3 Ultra	23

Claude Opus 5.5 led the same chart at 58. So ML4 moves open models from the US and Europe up a long way, but it still sits below the closed frontier and the best open models from China.

Pricing

Per 1M tokens	Launch list price	From October 29
Input	$1.36	$0.70
Cached input	$0.14	$0.07
Output	$4.18	$2.10

During the first weeks of the preview, Mistral ran a 50% launch sale at $0.68 / $2.09. When the weights shipped, it set the list price at roughly that level. There's no published batch rate for Large 4 yet. On OpenRouter, the EU-hosted route costs 1.1 times the global rate and has a 512K context window instead of 1M.

Open weights and the license

The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, with a separate NVFP4 checkpoint for Blackwell GPUs (B200, B300, GB200 and GB300). Mistral's serving examples use eight GPUs for FP8 and four to eight for NVFP4.

The license is the Mistral Ecosystem Open-Weight License, a modified MIT license. It allows use, modification, fine-tuning, redistribution and commercial use, including self-hosting on your own infrastructure. The catch is the model-as-a-service clause. If you or an affiliate run a business that sells access to models, you need a separate commercial license from Mistral. You can use ML4 freely inside your own company, but you can't sell API access to it without that license.

Built and served in Europe

Mistral is pitching ML4 to European buyers. It was trained entirely in Mistral's own European datacenters, and the API serves it from Europe, so data can stay in the EU. Banks, government agencies, healthcare providers and manufacturers that have to follow GDPR and data-residency rules can't get that from the US closed APIs.

Mistral raised €3 billion at a €21 billion valuation in September and is targeting 1 GW of European compute by 2030. On October 21, it announced a partnership with NVIDIA to apply ML4 to physics simulation and industrial design in aerospace, automotive and semiconductors, using CUDA-X, Warp and Omniverse, with open benchmarks and datasets planned but not yet published.

Should you use it?

ML4 makes the most sense if your data has to stay in the EU, or if you want a model this capable running entirely on your own hardware. Open weights also let you fine-tune it and move it between clouds without depending on one vendor. For security teams it's one of the strongest open models available, and it takes on legitimate vulnerability work that some closed models refuse.

If you only care about capability per dollar and data location doesn't matter, the case is weaker. DeepSeek V4.1 Flash scores about the same on the Artificial Analysis index for a fraction of the price, and GLM-5.3 and Kimi K3 score higher. Run ML4 and those models on your own tasks and compare the cost per finished task before you switch. Most benchmark numbers so far come from Mistral, and now that the weights are public, independent results should start appearing over the next few weeks.

Compare open models on A2Agent

A2Agent gives you DeepSeek, GLM, Kimi, MiniMax, Qwen and other models through one API key, so you can run the same tasks across them and measure cost and quality side by side before deciding where ML4 fits.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.a2agent.me/v1",
)

for model in ["deepseek-v4.1-flash-1017", "glm-5.3"]:
    response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": "Summarize the GDPR rules on data transfers outside the EU."}],
    )
    print(model, response.usage.total_tokens)
    print(response.choices[0].message.content)

See the Models page for current prices. New accounts get $5 in free credit.

Sources

• Introducing Mistral Large 4 (Mistral AI)
• Mistral Large 4 model docs (Mistral AI)
• Mistral Large 4 becomes top non-Chinese open model on Artificial Analysis Intelligence Index (OfficeChai)
• Mistral's blog says 49 billion active parameters, its docs say 52 billion (MIXED)
• Mistral Large 4 open weights: what changed since preview (BrainDetox)
• Mistral Large 4 price cut (TokenCost)
• Mistral and NVIDIA point Mistral Large 4 at industrial engineering simulation (AlphaSignal)