Mistral Large 4: a 1-trillion-parameter model with open weights. What it means for Pakistani teams
Mistral launched a preview of Mistral Large 4 on 6 Oct and says the weights come out this month. What we know, which claims are unverified, and why it matters here.
Mistral has unveiled a 1-trillion-parameter AI model, and it says the weights will be open. If your company wants to keep client data in-house, this one matters.
Paris-based Mistral launched a public preview of Mistral Large 4 (ML4) on Tuesday, 6 October 2026. According to Mistral’s announcement, developers can try it today through the preview API on Mistral Studio, and the company says it will release the model weights “by the end of the month”.
Here is what we know, which claims are still Mistral’s own, what we don’t know yet, and why it matters for Pakistani professionals.
First, what does “open weights” mean?
An AI model’s “weights” are the numbers it learned during training. When a company releases them, anyone allowed by the licence can download the model and run it on their own computers or cloud, instead of sending every request to the company’s servers. CNBC puts it simply: open models “can be modified and self-hosted”, unlike closed systems from OpenAI and Anthropic.
Open weights are not the same as “free for anything”. The licence decides what you may do with them (more on that below).
What Mistral announced
All of the following comes from Mistral’s announcement, unless another source is named.
| What | Detail |
|---|---|
| Name | Mistral Large 4, “unofficially ML4, very officially: le Chonk”, per Mistral |
| Size | 1 trillion parameters in total, with 49 billion active at a time |
| Input and output | “Natively multimodal”, says Mistral, so it reads images and documents as well as text. VentureBeat reports, citing co-founder Guillaume Lample, that it still produces text output only |
| Available now | A public preview through Mistral’s API |
| Preview price | Mistral’s announcement page lists $1.36 per million input tokens and $4.18 per million output tokens. VentureBeat says the company did not provide API pricing in the materials it reviewed, so check Mistral’s pricing page before you budget |
| Languages | Training data spanned “more than 160 languages, including every official language of the European Union”, according to Mistral |
| Focus areas | Coding, cybersecurity, finance, law, manufacturing, and reading images, charts and documents (Mistral, CNBC, VentureBeat) |
| Where it runs | Mistral says ML4 “will be able to run on private cloud or on-premise”, and that it will also be offered in multiple regions, including a European deployment Mistral runs itself |
When do the weights come out?
The two sources give slightly different wording, so here are both:
- Mistral says: “We will release the weights by the end of the month.”
- VentureBeat reports the company plans to publish the weights on 27 October, after a roughly three-week testing period with developers, cybersecurity leaders and government authorities.
Either way, the safe version is: later in October 2026, according to the company. Until then, Mistral says it is “red-teaming” (stress-testing) the model with cybersecurity leaders, vetted partners and state authorities.
VentureBeat also reports that the weights are expected under a custom Mistral licence. The licence text was not public in our sources.
The big claim: “strongest open-weight model outside China”
According to CNBC, Mistral said in a statement that ML4 is the strongest open-weight model developed outside China by a “substantial margin”, and that it will rank among the top open-weight models globally once released. Mistral’s own announcement says it significantly outperforms “any open-weight model developed in the US or Europe”.
CNBC adds its own context: the most capable open models so far have been Chinese, and the new model “still lags behind the frontier in areas such as coding”.
The benchmark scores: Mistral’s own early results
Treat every number in this section as Mistral’s claim, not an independent result. VentureBeat reports that, as of publication, ML4 did not yet appear in Artificial Analysis’ public evaluations or on the DeepSWE leaderboard, so Mistral’s ranking claim “remains provisional until outsiders can test the final model and released weights”. Mistral itself says the model is still being trained and will keep changing before the weights are released.
| Test | Mistral’s reported score | What VentureBeat found when it checked |
|---|---|---|
| DeepSWE v1.1 (long software-engineering tasks) | 61.7% | Competitive, especially against Western open models, but the live DeepSWE leaderboard’s best configurations put GLM-5.3 and Kimi K3 at about 69% and leading closed models around 74%. VentureBeat says it “does not establish an outright coding lead” |
| Harvey’s Legal Agent Benchmark | 15% task-pass rate | The competitor figures in Mistral’s chart match Vals.ai’s public leaderboard. If ML4’s score used the same method, it would lead those open rivals, but several closed models score higher |
| Finch (finance and accounting workflows) | 67% | VentureBeat could not find published Finch results matching the newer-model scores in Mistral’s chart |
| Cybench (40 security-competition exercises) | Solves 93% | Not independently checked in our sources |
Mistral also reports other results, including on cybersecurity and image tasks. We have left them out here because nobody outside Mistral had verified them at the time of writing.
What Mistral says about cybersecurity
Cybersecurity is the part Mistral pushes hardest. Its argument, as set out in the announcement and summarised by VentureBeat: closed AI providers’ safety filters can refuse legitimate defensive work, such as proving a software flaw is real, so security teams benefit from a strong model they control themselves.
In CNBC’s report, Lample said: “Further, the cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.”
Mistral also says that, despite the strong cyber scores, ML4 refuses malicious cybersecurity requests more often than other open models on the tests it ran. That is also Mistral’s own result.
What we still don’t know
- Independent test results. No outside evaluator had published scores for ML4 in our sources.
- The licence terms. VentureBeat says “a custom Mistral license”. What it allows for commercial use, and whether any company size or usage limits apply, is not yet public in our sources.
- The exact release date. Mistral says end of October; VentureBeat says 27 October.
- Urdu quality. Mistral says more than 160 languages, but our sources don’t list them or say anything about Urdu.
- Final scores. Mistral says the preview “continues to improve rapidly”, so the released model may score differently from the preview.
We will update this article when the weights and licence come out.
Why this matters in Pakistan
- Client data can stay in the building. Banks, hospitals, law firms and software houses with strict client contracts often can’t paste sensitive data into a foreign chatbot. A strong open-weight model that runs on your own servers, or a cloud you choose, may let you use AI on that work without the data leaving your control.
- Less dependence on one provider. If a company relies on one closed AI service and it changes prices, rules or access, work stops. Open weights give you a fallback you control. Mistral makes the same argument for security teams, where losing access “mid-incident” is itself a risk, it says.
- But a 1-trillion-parameter model is not a laptop model. Our rough maths: at about 1 byte per parameter (8-bit precision), 1 trillion parameters needs on the order of 1 TB of memory just to hold the weights, before anything else. Only 49 billion parameters are active at a time, which helps speed, but the whole model still has to be loaded. For most Pakistani teams this means a serious GPU server or a rented cloud machine, not an office PC. Smaller models built on ML4, which Mistral says are coming, may be more practical.
- Software houses and freelancers get a new option for client work. If a foreign client asks “where does our data go when you use AI?”, “on a server we control” is a stronger answer than “a chatbot”. Keep the licence in mind before you build a product on it.
- Where Pakistan’s own plans fit. For what the government has set out on AI, see our breakdown of the National AI Policy 2025 in 6 numbers.
Kaam ki baat: how to test ML4 before you switch
Open weights mean a bank, hospital or software house can run a strong AI model on its own machines, so client data never has to leave the building. Before you switch, wait for independent tests and try the preview on one real task from your work.
- Pick one real task. A code review, a contract summary, a spreadsheet clean-up, a scanned invoice. Use the same task for every model you compare.
- Use the preview with dummy data only. The preview runs on Mistral’s servers, so it is not “in-house” yet. Remove client names and sensitive details before testing.
- Compare side by side. Run the same prompt on the tool you use today and on ML4. Score the answers yourself: correct, partly correct, wrong.
- Test your languages. If your work involves Urdu or Roman Urdu, test that too. Mistral hasn’t said how well ML4 handles Urdu. (For Urdu tools that work today, see Urdu AI you can use today.)
- Read the licence when the weights land. Check commercial use, any limits, and what you must show users, before you build anything for a client.
- Wait for outside benchmarks. Mistral’s scores are its own early results. Give independent testers a few weeks after the weights are released.
- Do the hardware sums with your IT team. Self-hosting a model this big costs real money. Compare the server or cloud bill with API pricing for your actual usage.
FAQ
Is Mistral Large 4 free?
The preview is a paid API. Mistral’s announcement page lists $1.36 per million input tokens and $4.18 per million output tokens. When the weights are released, downloading them should not require paying per request, but you will need your own hardware, and the custom licence will decide what you may do.
When can I download it?
Mistral says “by the end of the month” (October 2026). VentureBeat reports a planned date of 27 October.
Is it really better than ChatGPT or Claude?
Mistral’s main claim is about open-weight models, not that it beats every closed model. VentureBeat notes that on the DeepSWE coding leaderboard, leading closed models score higher than Mistral’s reported result, and CNBC reports the model “still lags behind the frontier in areas such as coding”. All the scores are Mistral’s own early results for now.
Does it understand Urdu?
Mistral says its training data covered more than 160 languages, but our sources don’t say whether Urdu is one of them or how well it works. Test it yourself on your own Urdu or Roman Urdu text.
Why is it called “Le Chonk”?
According to VentureBeat, it nods to an online joke from June about an imaginary giant Mistral model called “Le Chaton Fat”. “Chonk” is internet slang for a very large cat.
Sources
- Mistral AI, Introducing Mistral Large 4 . Checked 6 Oct 2026 . Company announcement, 6 Oct 2026; archived copy saved 6:31 PM PKT and re-checked live 7:45 PM PKT
- CNBC, Mistral unveils new AI model it says rivals best open systems from China . Checked 6 Oct 2026 . By Kai Nicol-Schwarz. Published 6 Oct 2026, 9:00 AM EDT (6:00 PM PKT); archived copy saved
- VentureBeat, Mistral debuts Large 4 'Le Chonk', a 1-trillion parameter text output model with high benchmarks planned for open weights release . Checked 6 Oct 2026 . By Carl Franzen. Published 6 Oct 2026, 6:00 AM PT (6:00 PM PKT); archived copy saved
Drafted with AI assistance from the sources listed above; every figure and link is checked against those sources before publishing. Spotted an error? Tell us. Discuss it on LinkedIn.