Technology & Science
Mistral previews 1-trillion-parameter model, with weights due Oct. 27
The French startup is making its new model available through an API preview and says it will publish the weights later this month. Its benchmark comparisons remain company claims, and advanced cyber capabilities are limited to a red-team program.
Executive summary
- A sparse model at one trillion parameters: Mistral says Large 4 has one trillion total parameters, with 49 billion active for each token, and accepts text and image inputs. The company has opened an API preview and says the model's weights will be released on Oct. 27.[1][2]
- Performance claims await outside comparison: Mistral says Large 4 outperforms several open-weight and Chinese models, and cites evaluations by vals.ai on legal and financial tasks. Reuters reported the company did not specify the Chinese models behind its comparison; Le Monde noted independent aggregate rankings had yet to confirm the broad claim.[1][2][3]
- More capable cyber mode stays restricted for now: Mistral says a less-moderated configuration with expanded cybersecurity capabilities is limited to selected cyber experts, partners and authorities for real-world testing before any public release. The company says the controls are part of its red-team process.[1]
Mistral AI on Tuesday previewed Large 4, a multimodal model it describes as a mixture-of-experts system with one trillion parameters overall and 49 billion active for each token. The company says the model can process text and images and is available now through an API preview. It plans to release the model weights on Oct. 27; the weights are not yet public.[1][2]
Mistral said it trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs at its European data centers. That scale shows the computing resources behind this release, but the company has not disclosed a training-energy figure or a cost comparison that would establish how efficiently it used them.[1]
Mistral says Large 4 leads other open-weight models and several competitors from China on coding, reasoning, science and multilingual tasks. Reuters reported that Mistral chief executive Arthur Mensch did not name the Chinese models behind his comparison. Mistral also cites an evaluation by vals.ai on legal and financial tasks, but Le Monde noted independent aggregate rankings had yet to confirm the broad comparison. The evidence is a mix of company benchmarks and one specialist evaluator, not an established overall ranking.[1][2][3]
Mistral is positioning the release as a step toward competing with leading closed and open models. The API preview lets customers test the system before weights arrive, but public access to the weights is scheduled for Oct. 27, according to Reuters; Mistral's own post says the release is due by the end of the month.[1][4]
Mistral says a configuration with reduced moderation and expanded cyber capabilities is being made available only to cyber experts, selected partners and authorities for real-world red teaming. The company says the restricted program is intended to expose risks before the model weights are released. It has not announced a general release of that configuration.[1]
The approach leaves two release questions open: whether external evaluators will substantiate the performance comparisons, and what safeguards will accompany the weights when they become public. Neither answer is settled by the preview announcement.[1][2]