Mistral has published the weights on Hugging Face under a modified MIT license, allowing developers and businesses to run, modify, and commercially deploy the model under specified conditions. Organizations offering certain model-hosting services must obtain a separate commercial agreement.
The company is positioning Large 4 as an alternative for enterprises seeking greater control over their AI infrastructure. Because the model can operate on private servers, organizations can customize it for internal applications without relying entirely on externally hosted AI services.
That flexibility comes with substantial hardware requirements. According to deployment documentation, the standard model can run on eight Nvidia B300 GPUs, while a more compressed configuration requires four GB300 GPUs. Hosted access is also available through Mistral's API and platforms including Microsoft Foundry.
Large 4's most significant improvements are in coding efficiency and operating costs.
On Artificial Analysis' Coding Agent Index, the model scored 48 points, narrowly behind China's GLM-5.3 at 49. However, Large 4 completed the evaluated tasks using approximately 47% fewer tokens, suggesting it can deliver comparable coding results with less generated output.
Mistral's own testing showed additional improvements. Its Terminal-Bench 4.0 score increased from 28% during the preview to 41% after further training, while its reported FrontierCode 2.0 result reached 19.7%.
Those figures require some qualification. Independent testing by Artificial Analysis produced lower scores on both benchmarks under its own evaluation setup, highlighting how results can vary depending on the software and configurations used.
Large 4 also remains substantially behind the strongest proprietary coding systems. Anthropic's Claude Opus 5.5 scored 62 on the Coding Agent Index, compared with Large 4's 48.
The broader performance gap is similarly apparent in Artificial Analysis' updated Intelligence Index.
Large 4 achieved 26 points in its high-reasoning configuration, improving only slightly from the preview's 25. That makes it the highest-scoring model with publicly available weights developed outside China in the evaluation, but it remains behind several Chinese competitors.
DeepSeek V4.1 Flash and GLM-5.3 each scored 32 points, while MiMo-V2.6-Pro reached 30. Google's Gemini 4 Carbon led the overall rankings with 59 points, followed by Claude Opus 5.5 at 58.
Despite the modest improvement in overall intelligence, Large 4 delivered substantial efficiency gains. Artificial Analysis measured generation speed at 161 tokens per second, compared with 111 for the preview.
The average cost of completing an Intelligence Index task also declined approximately 63%, falling from $5.54 to $2.06. That makes Large 4 less expensive per evaluated task than GLM-5.3 at $5.01 and Kimi K3.1 Pro at $3.68, although DeepSeek V4.1 Flash and MiMo-V2.6-Pro remain considerably cheaper.
Mistral has also reduced API pricing to $0.70 per million input tokens and $2.10 per million output tokens, roughly half the preview's original rates. Cached input costs $0.07 per million tokens.
Cybersecurity is another area where Mistral has concentrated its training efforts.
Large 4 scored 51 on Artificial Analysis' Cyber Index, which evaluates a model's ability to identify and address software vulnerabilities while preserving existing functionality. That places it third among open-weight models in the company's comparison, behind DeepSeek V4.1 Flash at 57 and MiMo-V2.6-Pro at 56.
Mistral also reports that Large 4 rejects 96.6% of malicious cybersecurity requests, a higher refusal rate than competing open models in its evaluation. The company acknowledges that advanced cybersecurity capabilities can also be misused and plans to strengthen safeguards.
Select partners, including Orange Cyberdefense, are testing a specialized version intended for penetration testing. Access requires additional review and contractual restrictions.
Beyond technical performance, Mistral is emphasizing the commercial benefits of model ownership and customization.
Marjorie Janiewicz, Mistral's chief revenue officer, described customer demand for AI systems that organizations can adapt and maintain independently.
“I don’t want to get disconnected. I want to own my model,” Janiewicz said in an interview with VentureBeat. “Second, I want to customize it to ensure that I derive value. And then third, it’s business continuity.”
The licensing arrangement supports that approach for internal deployments and many commercial applications, but introduces restrictions for businesses that provide third parties with hosted model access. The requirement depends on the nature of the service rather than the company's revenue or size.
Mistral also emphasizes that Large 4 was trained entirely in Europe, positioning it as an option for organizations seeking greater control over where and how their AI systems operate.
The company intends to use Large 4 as the foundation for future specialized models, although it has not disclosed plans for smaller versions.
For now, the independent results establish a clear distinction between Large 4's capabilities and its competitive advantages. The model does not match the overall performance of leading Chinese open-weight systems or the strongest proprietary alternatives, but its coding results, faster inference, and reduced task costs make it a competitive option for organizations prioritizing efficiency and deployment control.
With its weights now publicly available, Large 4 gives enterprises another option for operating advanced AI models on their own infrastructure, even as Mistral continues working to narrow the performance gap with larger competitors.
This analysis is based on reporting from thedecoder & Venture Beat.
Image courtesy of Mistral.
This article was generated with AI assistance and reviewed for accuracy and quality.
