Mistral Large 4 is the French AI lab’s newest flagship model. Mistral built it to run on your infrastructure once it ships its weights. Mistral released it as a public preview on October 6, 2026. It has roughly 1 trillion total parameters in a sparse mixture-of-experts design. Developers and enterprise teams often look for self-hosted alternatives to closed models like GPT and Claude. Mistral Large 4 is the clearest attempt yet at a frontier-scale open-weight option from outside the US and China.
What Is Mistral Large 4
It is a natively multimodal model that accepts text and images and returns text. Mistral’s own announcement describes it as the lab’s “largest and most capable model to date.” The company calls it ML4.
They have not released the weights yet. The public preview runs only through the Mistral API and Mistral Studio. Mistral says it will ship open weights by the end of October 2026, after safety testing. Early coverage from The New Stack suggests the weights may use a custom license. That would differ from the Apache 2.0 terms Mistral applied to Large 3.
How Mistral Large 4’s Architecture Works
ML4 uses a sparse mixture-of-experts architecture with about 1 trillion total parameters. Mistral’s own figures put active parameters per token at roughly 52 billion, though some early third-party estimates cited 49 billion. The model also includes a 1.6-billion-parameter vision encoder for image understanding.
Mistral trained the model using about 4,000 Nvidia GPUs. Pierre Stock, Mistral’s VP of Science, described that figure as “two to three times less” compute than Chinese competitors use. Independent benchmarking firm Artificial Analysis measured a context window of roughly 512,000 tokens during early testing. Mistral’s own materials do not state an official figure.
Mistral Large 4 Pricing and Availability
Standard API pricing runs $1.36 per million input tokens and $4.18 per million output tokens. Cached input tokens cost $0.14 per million. Mistral is offering a 50% launch discount for the first two weeks. That drops the rate to $0.68 per million input tokens and $2.09 per million output tokens.
Mistral serves the model from its own European data centers today. The company plans wider regional availability once the open weights ship. Running the full model locally at 4-bit quantization needs roughly 500GB of memory. That exceeds what a single high-end consumer GPU like the RTX 5090 can hold. That makes self-hosting a realistic option for enterprise infrastructure, not home hardware.
How Mistral Large 4 Performs on Benchmarks
On Mistral’s own numbers, ML4 scores 61.7% on the DeepSWE v1.1 coding benchmark. It also scores 93% on Cybench, a cybersecurity capture-the-flag test. A blind human evaluation run by Surge AI ranked ML4 second of five models, at 3.74 out of 5. Claude Opus 5 led at 4.22.
Independent results tell a more mixed story. Artificial Analysis scored Mistral Large 4 at 38 on its Intelligence Index, tying OpenAI’s GPT-6 Luna. Several Chinese open-weight models scored higher on the same index. GLM-5.3 scored 45, Kimi K3 scored 44, and Xiaomi’s MiMo-V2.6-Pro scored 46.
Mistral Large 4 performed better on security-specific tests. It scored 50 on Artificial Analysis’s Cyber Index, ahead of both GPT-6 Astra (33) and Claude Opus 5.5 (29). Those two closed models declined some cybersecurity tasks on safety grounds, which pulled their scores down.
DevX has covered how generative AI is reshaping software development. More vendors now pursue agentic coding and security use cases like this.
Cost efficiency is where Mistral Large 4 lags furthest behind. Artificial Analysis calculated $1.13 per Intelligence Index task at list price, or $0.57 with the launch discount. GPT-6 Luna costs $0.07 per equivalent task, and most Chinese open-weight rivals land between $0.13 and $0.27.
Why Mistral Large 4 Matters for Developers and Enterprises
The real story is not raw benchmark scores. It is who gets to run a frontier-scale model on their own terms. French President Emmanuel Macron called the release “a third way in AI.” That framing positions Mistral as a counterweight to US and Chinese labs.
Mistral is targeting specific enterprise niches rather than general chatbot use. Mistral’s own announcement lists cybersecurity, finance, legal work, scientific computing, and engineering as priority use cases.
Pierre Stock said Mistral wants to work with “trusted partners and governments.” That would let trusted partners and governments use the weights defensively once they ship.
For teams weighing self-hosted inference against closed APIs, infrastructure cost matters as much as model quality. DevX’s rundown of cloud computing types covers the tradeoffs between on-premises, private, and public cloud deployment. Those tradeoffs apply directly to a model this size.
EU-based teams with data residency requirements have good reason to wait for ML4’s weights. That beats defaulting to a US-hosted API by default.
Common Mistakes and Tradeoffs to Watch
The biggest mistake is assuming ML4 beats Claude, GPT, or Gemini on quality. By Mistral’s own human evaluation, it does not. Its value is openness and self-hosting potential, not a benchmark crown.
The second mistake is budgeting around launch pricing as permanent. The 50% discount expires two weeks after release, doubling the effective cost per task.
The third mistake is planning a self-hosting rollout before the weights and license actually ship. Nothing is downloadable yet, and Mistral has not confirmed the license terms.
Key Takeaways
- Mistral released Large 4 as a public preview on October 6, 2026. It has about 1 trillion total parameters and roughly 52 billion active per token.
- Mistral expects to release open weights by the end of October 2026, under a license it has not yet confirmed.
- Standard pricing is $1.36 per million input tokens and $4.18 per million output tokens. A 50% discount applies for the first two weeks.
- Artificial Analysis scored it 38 on its Intelligence Index, behind several Chinese open-weight rivals like GLM-5.3 and MiMo-V2.6-Pro.
- Self-hosting the full model requires roughly 500GB of memory, ruling out consumer GPUs and pointing to enterprise infrastructure.
Frequently Asked Questions About Mistral Large 4
What Is Mistral Large 4?
Mistral Large 4 is Mistral AI’s newest flagship model. It is a natively multimodal system with about 1 trillion total parameters, built on a sparse mixture-of-experts architecture.
Is Mistral Large 4 Open Source?
Not yet. Mistral released it as a closed public preview on October 6, 2026. The company says it will release open weights by the end of October.
How Much Does Mistral Large 4 Cost?
Standard pricing is $1.36 per million input tokens and $4.18 per million output tokens. Cached input costs $0.14 per million. A launch discount halves both rates for two weeks.
How Does Mistral Large 4 Compare to GPT-6 and Claude?
It ties GPT-6 Luna on Artificial Analysis’s Intelligence Index and trails Claude Opus 5 on human-rated coding quality. It leads both GPT-6 Astra and Claude Opus 5.5 on Artificial Analysis’s Cyber Index.
Can I Self-Host Mistral Large 4?
Only once Mistral releases the weights. Running it locally at 4-bit quantization needs roughly 500GB of memory. That requires enterprise-grade infrastructure, not a single consumer GPU.
What Is Mistral Large 4 Best Used For?
Mistral positions it for cybersecurity work, agentic coding, and analysis of financial or legal documents. It also fits scientific computing, especially for teams that need self-hosted or EU-based deployment.
Final Thoughts
Mistral Large 4 is not the model to pick if you only want the highest benchmark score. It is the model to watch if your team needs a frontier-scale system it can eventually self-host. That matters if you also need to meet your own compliance rules. Wait for the actual weights and license terms before planning a rollout. Test it against your specific workload once it ships.
Photo by Markus Winkler: Pexels
Noah Nguyen is a multi-talented developer who brings a unique perspective to his craft. Initially a creative writing professor, he turned to Dev work for the ability to work remotely. He now lives in Seattle, spending time hiking and drinking craft beer with his fiancee.
