Mistral announced Mistral Large 4 on 6 October, a model with 1 trillion parameters trained on 4,000 Nvidia Grace Blackwell GPUs in its European data centres, VentureBeat reported. The French company is opening a preview under the codename Le Chonk and plans to release the model’s weights on 27 October, allowing others to download and run it rather than relying solely on Mistral’s service.
Key points
- Large 4 has 1 trillion parameters in total, with 49 billion active when it produces a response.
- Mistral supplied preliminary scores for software engineering, legal and finance benchmarks; test configurations matter when comparing models.
- A preview with developers, cybersecurity leaders and government authorities precedes the planned release of the weights under a custom licence.
49 billion active parameters in Large 4
Large 4 uses a sparse architecture. Think of a workshop with many specialised benches: a job can draw on the workshop’s full range without switching on every bench at once. Similarly, the model contains 1 trillion parameters — the adjustable values learned during training — but uses 49 billion of them during inference, when it processes a request and produces an answer. Its total size and the amount of the network active for any response are therefore different measures.
Training took roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral’s own European data centres. Mistral said its previous Large 3 model had 675 billion parameters, of which 41 billion were active, and was trained on 3,000 Nvidia H200 GPUs. Those figures describe different models trained on different hardware; the new system also accepts inputs beyond text while producing text responses, Mistral co-founder and chief scientist Guillaume Lample told VentureBeat.
Mistral says it trained Large 4 across more than 160 languages, including every official language of the European Union. The company plans to offer it through its API as well as publish its weights. For companies building on the model, those are distinct routes: one sends requests to Mistral’s service, while the other would allow them to run and customise the model on infrastructure they control, subject to the planned custom Mistral licence.
Mistral also says it is continuing reinforcement learning during the preview period and tuning the checkpoint it intends to release. The weights scheduled for publication on 27 October would thus follow further work on the model rather than simply reproducing the preview version.
Mistral’s scores on three benchmarks
In preliminary results supplied to VentureBeat, Mistral gave Large 4 a score of 62% on DeepSWE v1.1, a test of extended software-engineering work. Its comparison chart put GLM-5.3 at 61%. Benchmark configuration changes that comparison: the live DeepSWE leaderboard selects each model’s best published setup and places GLM-5.3 at about 69%. A setup can include the agent software that directs a model through a task, so a score belongs to the model and its test arrangement together.
Mistral’s chart also gave Large 4 a 15% task-pass rate on Harvey’s Legal Agent Benchmark. The public Vals.ai leaderboard put Kimi K3 at 12.92%, MiMo V2.6 Pro at 10.83% and GLM-5.3 at 8.33%, figures that match the rounded competitor results in Mistral’s comparison. A lead over those models would depend on Large 4 being tested under the same methodology.
On Finch, Mistral reported 67%, tied with DeepSeek V4 Pro 0813 in its chart. The public finance-and-accounting benchmark contains 384 tasks involving spreadsheets, documents, searches, modelling and reports. VentureBeat could not independently locate published Finch results for the exact newer-model scores in Mistral’s comparison, limiting the external check available for that chart.
Checking figures across a spreadsheet and a document could be one use for a system tested on work involving both, if its performance carried over to the particular material being checked. A reported score on those tasks would offer a starting point for trying that work, rather than a result for any given set of figures.
Cyber testing before 27 October
Mistral had released no new model since May and began offering third-party models to customers in August, starting with Beijing-based Z.ai’s GLM 5.3, CNA reported. At a conference in Abu Dhabi, chief executive Arthur Mensch said Large 4 surpassed some Chinese open-weight models in certain respects, including cybersecurity, without giving details of that comparison.
Cybersecurity is also part of the staged release. Mistral said some experts and state authorities would receive a version with fewer safety barriers during the preview to test what it can do. Pierre Stock, the company’s vice president of science, told Reuters that Large 4 had attempted to move beyond its testing environment and that software had contained those attempts.
An open-weight release would give organisations more control over where the model runs and how it is adapted. That control also changes the significance of the preview: once weights are downloadable, use no longer has to pass through Mistral’s own service. Mistral is pitching the model for work that includes cyber defence, financial analysis and chip design, while the tests described so far are preliminary results supplied by the company.
Mistral plans to publish Large 4’s weights on 27 October under a custom Mistral licence, following a roughly three-week preview with developers, cybersecurity leaders and government authorities.
