TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Mistral has launched a public preview API for Mistral Large 4, a natively multimodal model it describes as having 1 trillion total parameters and 52 billion active parameters. The company says it will release the model weights by the end of the month, after further red-teaming and publication of additional technical details.
Mistral has launched a public preview of Mistral Large 4, a 1-trillion-parameter natively multimodal model with 52 billion active parameters, and says it plans to publish the model weights by the end of the month. The preview is available through Mistral Studio; the planned weight release would let organizations run and adapt the model themselves rather than access it only through Mistral’s hosted service.
Mistral describes Large 4 as its largest and most capable model to date, while acknowledging that it is still refining the system. The company says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters, and that the preview runs on that same infrastructure. It says the model will be offered across multiple regions, including a European deployment operated by Mistral under European law.
The company reports results across several areas, including coding, agentic workflows, multimodal understanding, and cybersecurity. On the Artificial Analysis Cyber Index, Mistral says Large 4 ranks among the top five models globally and leads open-weight models developed outside China. It reports an 82% score on an index test involving reproducing and patching a vulnerability, and a 93% score on Cybench, a set of 40 security exercises. These results are claims reported by Mistral; the supplied announcement does not include independent verification details for every comparison.
For software tasks, Mistral reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. It says its combined Coding Agent Index score is 49.8%. In a blind human review of coding output conducted with Surge AI, the company says Large 4 Preview ranked second among five models, receiving an average rating of 3.74 on a five-point scale. Mistral also reports a 59.9% result on AutomationBench, which tests 657 business workflows across applications including Gmail, Google Sheets, Slack and Salesforce.
Open Weights for Enterprise Control
The planned weight release could matter to companies that need to run AI models in private environments, tune them for specific work, or maintain access under their own security and compliance rules. Mistral argues this is especially relevant in cybersecurity: it says provider-level refusals can interfere with legitimate vulnerability research and incident response, while self-deployment gives organizations more control over how security work is handled.
That flexibility also brings responsibilities. A model capable of assisting with vulnerability reproduction and other security tasks could be used in both defensive and harmful ways. Mistral says it is red-teaming Large 4 in real-world settings with cybersecurity leaders, vetted partners and state authorities before releasing the weights. The announcement does not provide a public account of the testing methods or the safeguards that will accompany the release.
The development also reflects a wider competition over access to advanced AI infrastructure and deployment. Mistral presents European-operated hosting and customer-controlled weights as options for organizations seeking alternatives to relying solely on external services. Whether the model’s reported performance and deployment options meet enterprise needs will depend on independent testing, technical documentation, cost, and the conditions attached to use.
As an affiliate, we earn on qualifying purchases.
Preview Before the Weight Release
Large 4 is available now as a hosted API preview, but that is not the same as public access to its weights. Mistral says the weights are due by the end of the month and that it will publish more information on the model architecture, additional benchmarks and post-training methods before or around that release. Until then, outside developers cannot independently inspect or deploy the weights based on the announcement alone.
Mistral says a significant share of the training data was multilingual and covered more than 160 languages, including every official language of the European Union. It also says the model was developed using the same training, customization and reinforcement-learning environment it offers customers through Mistral Forge. The announcement describes these choices as part of a broader effort to build models and infrastructure in Europe, but does not provide a full account of the training data or model architecture.
The company says Large 4 will serve as a foundation for a new generation of specialized and optimized Mistral models. That makes the preview both a product launch and an early indication of the company’s planned model lineup. No release schedule or specific capabilities for those future models was included in the supplied announcement.
“Today, we’re launching a public preview of Mistral Large 4.”
— Mistral
As an affiliate, we earn on qualifying purchases.
Benchmarks and Release Details Pending
Independent confirmation remains limited in the announcement. Mistral provides benchmark names and scores, but the material supplied does not include full evaluation reports, test conditions, or independent verification for all comparisons. Results can depend on prompts, model versions, tools and scoring methods, so the figures should be read as company-reported results until further documentation is available.
The exact date of the weight release, the license and usage terms, and which model components will be included have not been specified. Mistral also has not yet published the additional architecture and post-training information it says is forthcoming. It is unclear what moderation or safeguards will apply to the released weights or how the model’s cyber capabilities will be governed after organizations deploy it themselves.
The source material ends partway through its description of agentic workflow benchmark comparisons. It therefore does not establish the complete field ranking for AutomationBench or provide a full set of results across all the model’s claimed capabilities.
As an affiliate, we earn on qualifying purchases.
Technical Details Before Public Weights
Mistral says it will continue real-world red-teaming with cybersecurity leaders, vetted partners and state authorities as it works toward the planned weight release. The company has not named a specific release date beyond saying the weights are expected by month’s end.
It also says it plans to share more on the architecture, benchmarks and post-training methodology. Those materials, along with the release terms and independent evaluations, should help developers and organizations judge whether the model’s reported results translate to their own workloads. For now, users can test the hosted preview through Mistral Studio, while self-hosting remains dependent on the promised weight release.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Mistral Large 4?
Mistral Large 4 is a natively multimodal model that Mistral describes as having 1 trillion total parameters and 52 billion active parameters. The company says it is its largest and most capable model so far.
Can developers use it now?
Yes. Mistral says a public preview API is available through Mistral Studio. The model weights have not yet been released, so the current access described in the announcement is hosted access.
When will Mistral release the weights?
Mistral says it plans to release the weights by the end of the month. The announcement does not specify an exact date or provide the release license and usage terms.
How strong is Large 4 on cybersecurity tasks?
Mistral reports an 82% score on one Artificial Analysis Cyber Index test involving vulnerability reproduction and patching, and 93% on Cybench. These are company-reported results in the supplied announcement; full evaluation details and independent verification are not provided there.
What remains unknown about the release?
The exact release date, license, included model components and post-release safeguards are not yet clear. Mistral says it will publish more on the architecture, benchmarks and post-training methods, but those details are still pending.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
