On 6 October 2026, Mistral AI announced its largest model to date and promised to release the weights before the month is out. The nickname: "Le Chonk". For companies that want AI under their own control it's the news of the autumn, but the details decide whether it changes anything for you.
Mistral Large 4 is an open-weight language model from the French company Mistral AI with 1.05 trillion parameters, 52 billion of them active per token, that reads text and images and has a 1 million token context window. It has been available as a preview in Mistral's API since 6 October. The weights, which let you run the model on your own hardware, follow at the end of October according to Mistral, under a licence that hasn't been published yet.
What exactly is Mistral Large 4?
Mistral Large 4 is a mixture-of-experts model: it consists of many sub-networks, and only a small share of them computes each token. That makes it about as fast per answer as a 52-billion-parameter model while carrying the knowledge of one with over a trillion. The model page in Mistral's documentation lists "52B active parameters and 1.05T total parameters, and a 1.6B vision encoder", version v26.10, status "Public Preview".
According to Mistral's announcement, the model was trained "on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe". Mistral doesn't say which country or site. A European maker, trained in European data centres: for companies steering on digital sovereignty, that's the real selling point, more than any benchmark.
There's a Dutch angle too. ASML led the €1.7 billion funding round in September 2025 that valued Mistral at €11.7 billion, according to Mistral. ASML put in €1.3 billion for a stake of around 11%.
| Feature | Mistral Large 4 |
|---|---|
| Parameters | 1.05 trillion total, 52 billion active per token |
| Input | text and images |
| Context window | 1M tokens per Mistral, about 524K in the current preview API |
| API price per 1M tokens | $1.36 input, $4.18 output (temporarily $0.68 / $2.09) |
| Processing in Europe | yes, through Mistral's regional endpoints |
| Open weights | announced for end of October 2026 |
| Licence | not yet published |
The context gap isn't a typo. The documentation says 1 million tokens, but Artificial Analysis measures about half that in the preview API. Plan on 500,000 tokens for now, which is still plenty for almost any business document.
How good is Mistral Large 4 on independent measurements?
On independent measurements, Mistral Large 4 is a strong open model that sits clearly behind the best closed models. That's a soberer picture than the launch painted.
Mistral says the model scores "the highest of any model" on a cybersecurity test and that, through the evaluator Vals.ai, it beats OpenAI's GPT-6 Astra on legal and financial tasks. Vals.ai's own pages show a mixed picture. On the finance agent test, Mistral Large 4 scores 54.68% against 53.54% for GPT-6 Astra, a gap inside the error margin of just over 2 points. On the legal research test it trails at 31.73% to 39.42%, and on the overall Vals Index Mistral Large 4 ranks 34th of 49 while GPT-6 Astra ranks 7th. Mistral doesn't say which legal test it means, so the claim may concern a different task. But this is exactly the pattern to watch for: a launch picks the sub-tests the model wins.
Artificial Analysis, which tests all models under the same settings, gives Mistral Large 4 a 38 on its Intelligence Index, 66th of 217. The best closed models score between 52 and 58 there. Among open models it looks like this:
| Open model | Maker | Parameters (active) | Licence | Index score |
|---|---|---|---|---|
| MiMo-V2.6-Pro | Xiaomi (China) | not stated | MIT | 46 |
| GLM-5.3 | Zhipu (China) | not stated | own GLM licence | 45 |
| Kimi K3 | Moonshot (China) | 2.8T (104B) | own Kimi licence | 44 |
| DeepSeek V4.1 Flash | DeepSeek (China) | 552B (16B) | MIT | 39 |
| Muse Spark Open | Meta (US) | 816B (42B) | Muse Model License | 39 |
| Mistral Large 4 preview | Mistral (France) | 1.05T (52B) | not yet known | 38 |
Scores from Artificial Analysis, Intelligence Index v4.3.2, retrieved 17 October 2026. Mistral's claim to be the best open model "from US or Europe" roughly holds on this ranking: Meta's Muse Spark Open scores one point higher. The Chinese models lead.
Two caveats. The index is English-only and says nothing about Dutch; Mistral Large 4 isn't on EuroEval's Dutch leaderboard yet either. And a preview is a preview: the final release may score differently. How to read these numbers and what to test yourself is in our piece on evaluating AI benchmarks.
What does it take to run Mistral Large 4 yourself?
Running Mistral Large 4 yourself takes at least one server with eight top-tier GPUs, and then only with a compressed version of the model. Mistral hasn't published hardware requirements yet, so the arithmetic below is indicative.
Memory is the constraint. Every parameter takes memory, depending on the precision you load the model in:
- BF16 (2 bytes per parameter): 1.05 trillion × 2 = 2,100 GB
- FP8 (1 byte): 1,050 GB
- 4-bit (half a byte): 525 GB
On top comes memory for the context of running conversations, the KV cache. An NVIDIA H200 has 141 GB of memory, so an eight-GPU server has 1,128 GB. In FP8 the weights alone take 93% of that, leaving 78 GB for context. That doesn't work for real use. Two routes are realistic: FP8 split across two servers (16 GPUs), or a 4-bit version on one server.
The predecessor and a competitor make the estimate concrete. Mistral Large 3, at 675 billion parameters, runs in FP8 on "one 8xH200 node" according to its model card. Meta specifies at least 700 GB of GPU memory for Muse Spark Open, at 816 billion parameters, in 4-bit. Scale that to 1.05 trillion parameters and 4-bit lands around 900 GB. One 8×H200 server, with little room to spare.
For most companies of 50 to 500 people, that investment doesn't match the task. For document analysis, internal search or customer service in Dutch, a model of 100 to 130 billion parameters, such as Mistral Small 4 on four H100s, is almost always the better choice. Which hardware fits which model size is worked out in AI hardware for business.
Time saved
Save 10 hours per week on manually searching and summarising confidential case files
Which licence will the weights come under?
The licence for Mistral Large 4 is still unknown as of 17 October 2026, and for a company it's the biggest open question. Mistral uses two kinds of licence, and the difference hits some of our readers directly.
Mistral Large 3 and Mistral Small 4 are under Apache 2.0: free to use, modify and deploy commercially. Mistral Medium 3.5 is under a "Modified MIT License" that withdraws all rights if your company's global consolidated monthly revenue exceeds $20 million. That's about $240 million a year. Most companies with 50 to 500 staff are below it, but the test covers the whole group's revenue. A Dutch subsidiary of an international group can easily fall outside it.
Trade press reports a "custom Mistral license" for Large 4. Mistral hasn't confirmed that. Read the licence before ordering hardware, and look specifically for revenue thresholds, user counts and the terms for offering the model as a service. How the licences of all the major open models compare is set out in open-weight licences compared.
What does the AI Act mean if you use Mistral Large 4?
If you use Mistral Large 4 in your own systems, under the AI Act you're a deployer, not the provider of the model. The obligations for providers of general-purpose AI models sit with Mistral.
That only changes if you modify the model substantially. The European Commission's guidelines put the line at a modification using more than a third of the original training compute. No company fine-tune of a model this size gets there. The Commission itself writes that "currently few modifications may meet the criterion".
For Mistral, one more thing counts. Article 53(2) of the AI Act exempts providers of open models from part of the documentation duty, but "shall not apply to general-purpose AI models with systemic risks", meaning models trained with more than 10^25 operations. Mistral hasn't disclosed the training compute of Large 4. It is on the list of signatories of the EU code of practice for these models. What matters for you is the consequence: if you build your own AI system on the model, Mistral owes you technical documentation about it whenever the exemption doesn't apply. That's the case with systemic risk, but under the same guidelines also when the licence doesn't count as free and open, for example through "usage restrictions triggered by user scale thresholds" or a revenue threshold.
What should you do with Mistral Large 4 now?
What you do with Mistral Large 4 now depends on why you're looking at it. Most companies don't need to wait for the weights: the decision you can make today is whether the model beats what you already use on your own test set.
| Your situation | What to do now |
|---|---|
| You want a European provider for sensitive data, but no hardware of your own | Test the API with European processing on your own tasks, and compare the cost per task with what you pay now |
| You want a large model under your own control | Wait for the licence, cost out the hardware, and compare with the open models already available |
| You want the highest quality, whatever the origin | The closed frontier models score 14 to 20 points higher; Large 4 doesn't win there |
| You already run a smaller open model | Stay with it unless your test set shows a clear gap; a trillion-parameter model isn't an upgrade you make lightly |
For the wider question of how open models compare to the most expensive closed ones, see our analysis of open-weight models versus the frontier. The role of the Dutch GPT-NL alongside Mistral is covered in GPT-NL and European AI models, and the full picture from model choice to operations is in our guide to local AI for business.
Want to know whether Mistral Large 4, or a smaller open model, pays off under your own control for your use case? We work that out with you as part of local AI.