senn-techMistral Large 4: 1 trillion parameters, weights due end of October
On 6 October 2026 Mistral opened the public preview of Mistral Large 4, ML4 for short, nicknamed "le Chonk". It has 1 trillion parameters and is natively multimodal. The preview runs through the API in Mistral Studio; the weights are due by the end of the month. For an open-weight model the date is the actual news: until then nobody can run it or measure it independently.
What Mistral claims
According to Mistral, the model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe. The training data covers more than 160 languages, including every official EU language. Mistral puts the active parameters at 52 billion per token, while The Decoder says 49 billion, so we cannot tell which number is right.
A large part of the announcement is about security work. On one test from the Artificial Analysis Cyber Index, where a model has to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82 percent. Mistral says Claude Opus 5.5 and GPT-6 Astra score close to zero. Per The Decoder, the cause is the models refusing the task and points out that the test therefore also measures vendor policy as well as model ability. Mistral does not explain how ML4 tells legitimate vulnerability research from attack preparation. During the preview it is also red-teaming a variant with reduced moderation and extended cyber capabilities together with security firms, vetted partners and government agencies.
On coding, ML4 reaches 49.8 percent on the Artificial Analysis Coding Agent Index, ahead of DeepSeek V4 Pro and Qwen3.8 Max. In a blind rating by Surge AI it scores 3.74 out of 5, second of five models, behind Claude Opus 5 at 4.22.
What an independent party measures
The Artificial Analysis Intelligence Index combines ten benchmarks. For ML4 it reports 38. The predecessor Mistral Large 3 scored 9 and Mistral Medium 3.5 scored 14. Claude Opus 5.5 (Max) leads with 58.
Mistral has caught up with its own predecessor by a wide margin. "Strongest open model from Europe and the US" is Mistral's own statement, and it depends on which models Mistral chose to compare against.
Two meanings of "parameters" for hardware
In a mixture-of-experts model the active parameters set the compute per token, while the total sets the memory. They have to be counted separately. A trillion parameters take about 1 TB at 8 bit and about 500 GB at 4 bit; that is plain multiplication and no measurement. Our production machine for local models has four RTX 5090s with 128 GB of VRAM between them. ML4 does not fit on those cards alone, even heavily quantized.
That leaves offloading experts to system memory. With only 52 billion parameters active per token it works in principle. Speed then depends on the host's memory bandwidth and the PCIe bus, and how well the common inference servers handle a model of this size and design will only show once the weights are out.
What we do until then
We do not test what does not exist. When the weights appear, two questions come first: which license they ship under, since the announcement names none, and whether a quantization exists that fits a machine with offloading. Only after that is a comparison with the lane we run today worth the time. Anyone using the API preview sends prompts to Mistral. The announcement does not say where the preview is hosted. For data that we only process locally, that is not an option.
The security numbers deserve extra caution. A test that competitors fail because of their safety filters says little about how well ML4 finds vulnerabilities. Whether it actually beats the competition can be judged only with the weights and our own test setup.
Further reading
- Introducing Mistral Large 4 (Mistral, 6 October 2026): primary source with architecture, benchmarks and schedule
- Mistral Large 4 soll das stärkste offene KI-Modell aus Europa und den USA sein (The Decoder, German): classification with the Artificial Analysis values
- Simple Icons: mistralai.svg: logo, CC0
Are the Mistral Large 4 weights available yet?+
No. Mistral opened a public preview through the API in Mistral Studio on 6 October 2026 and says the weights follow by the end of the month. The announcement names no license. Until the release, Artificial Analysis still lists the model as proprietary.
How many parameters are active in Mistral Large 4?+
Mistral states 1 trillion parameters in total, 52 billion of them active. The Decoder's report says 49 billion. We use the figure from Mistral's own page and note that the two sources disagree.
Will Mistral Large 4 run on a server with four RTX 5090s?+
Not on the cards alone, by our arithmetic. Four RTX 5090s add up to 128 GB of VRAM. The weights of a trillion-parameter model take about 500 GB at 4 bit and about 1 TB at 8 bit. Offloading experts to system memory would make it possible, but slow. We have measured nothing, because the weights do not exist yet.