One trillion parameters. Mistral says it trained Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Europe; Quartz reports a different figure: 4,000. On Tuesday, the company launched the public preview of Mistral Large 4, nicknamed “Le Chonk.” Quartz reports Mistral’s claim that the model leads open-weight models outside China by a substantial margin.
TL;DR: Mistral launched the public preview of Mistral Large 4 (Le Chonk), a 1.05-trillion-parameter multimodal mixture-of-experts model with 49B active parameters and a 1M context window. Mistral reports 3,800 NVIDIA Grace Blackwell GPUs, while Quartz reports a different figure: 4,000. Open weights are planned for October 27. Quartz reports Mistral’s claim that ML4 leads open-weight models outside China by a substantial margin. Mistral highlights cybersecurity among its focus areas.
What Is Mistral Large 4 and Why the ‘Le Chonk’ Nickname?
Mistral Large 4 (ML4) is the French company’s biggest model yet: a one-trillion-parameter, natively multimodal system trained on its own European infrastructure. It launched as a public preview through Mistral’s API, with a full open-weight release planned for October 27, according to MarketScreener. The company lists cyber, coding, manufacturing, finance, and multimodal work among its focus areas, as CNBC reports.
The nickname is deliberate. “Le Chonk” - a playful take on internet slang for something large and heavy - nods to the community that had been rooting for Mistral to build a massive frontier model, VentureBeat reports. It is an unusual name for a flagship release. For earlier context on the company’s product direction, see our Mistral AI Summit recap.
Sources describe the release as a statement of intent: a European lab openly challenging both closed US frontier models and Chinese open-weight leaders. Independent benchmarks remain thin so far, with only Mistral’s own tests and limited third-party evals available at launch.
How Big Is the Model Under the Hood?
Mistral describes Large 4 as a 1-trillion-parameter multimodal model with 49 billion active parameters; MarkTechPost reports the more specific total of 1.05 trillion. The model uses a mixture-of-experts (MoE) architecture, so only a subset of the network fires for any given token.
Here is what sources confirm about the spec sheet:
- Total parameters: 1.05 trillion
- Active parameters per inference: 49 billion (MoE architecture)
- Context window: 1 million tokens, per Mistral’s model page
- Modality: natively multimodal
- Training infrastructure: 3,800 NVIDIA Grace Blackwell GPUs in Europe, per Mistral (Quartz reports 4,000)
- Availability: public API preview now, open weights on October 27
- Key claimed strengths: cybersecurity, coding, manufacturing, finance, multimodal reasoning
- Positioning: Quartz reports Mistral’s claim that ML4 leads open-weight models outside China by a substantial margin
A 1M-token context window means the model can process very long documents in a single pass. Combined with multimodal reasoning, Mistral is clearly targeting enterprise workloads — think legal, financial, and engineering documents rather than short chat exchanges.
| Spec | Mistral Large 4 (Le Chonk) |
|---|---|
| Total parameters | 1.05T |
| Active parameters | 49B |
| Architecture | Multimodal MoE |
| Context window | 1M tokens |
| Training hardware | 3,800 NVIDIA Grace Blackwell GPUs (Mistral); 4,000 reported by Quartz |
| Open weights | October 27 |
Why Is Mistral Releasing the Weights Only on October 27?
The public preview and the planned open-weight release are two stages. Mistral launched the public preview through its API; MarketScreener reports an October 27 date for publishing the trained parameters.
Why stagger it? Sources do not spell out Mistral’s reasoning, but the pattern is a familiar one for frontier open-weight releases: run a preview period first, then publish weights once the model has been exercised by early users. The company says it has narrowed the gap with the most capable frontier systems, and the open-weight step is how it proves that claim publicly — independent developers will be able to download, benchmark, and stress-test Le Chonk themselves.
The open-weight commitment is central to the launch. VentureBeat notes that benchmark results are planned for the weights release, putting independent community testing alongside Mistral’s own claims.
Can Mistral Large 4 Really Rival Chinese Open-Weight Leaders?
Quartz reports that Mistral claims ML4 leads open-weight models outside China by a substantial margin, framing the competition against Chinese open systems as well as closed US frontier models. CNBC describes the model as a rival to the best open systems from China.
The evidence so far is mixed in its independence. In Mistral’s own tests, Large 4 beats GLM 5.2, while Trending Topics notes that independent benchmarks were still missing at launch. The Decoder reports that the model makes a big leap forward in the independent Intelligence Index, a more encouraging data point for Mistral’s claims.
Coverage across sources converges on a few themes:
- Quartz reports Mistral’s claim that ML4 leads open-weight models outside China by a substantial margin
- The model aims to leapfrog both closed and open rivals, per TechCrunch’s framing
- Beating GLM 5.2 is based on Mistral’s internal testing
- Independent Intelligence Index results show a significant jump
- French and European press frame it as a challenge to US frontier models too, with Clubic comparing it against GPT-6 Astra
The real test arrives after October 27, when the weights go public and the broader benchmarking community can weigh in.
Why Does Mistral Push Cybersecurity as a Selling Point?
Cybersecurity is the headline capability, and sources single it out repeatedly. Coverage of the launch describes cyber capabilities as the selling point of ML4’s preview, alongside coding and multimodal reasoning. Mistral lists cybersecurity first among areas where the model is particularly effective, ahead of coding, manufacturing, finance, and multimodal work.
The strategic logic is visible in how the story is being told. If Mistral is racing China on open-weight AI — as one report puts it — then differentiation matters more than raw benchmark scores alone. A model marketed specifically for security work gives European enterprises and governments a reason to prefer it over rivals, particularly given that Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in Europe; Quartz reports 4,000.
That framing matters commercially. Mistral presents European infrastructure and cybersecurity as differentiators for organizations evaluating an open-weight model. Whether that pitch changes adoption or rankings will require customer evidence and independent testing.
How Does It Compare With GPT-6 Astra and US Closed Models?
Mistral positions Large 4 as Europe’s direct answer to the leading US frontier models, with GPT-6 Astra named explicitly among the rivals. Coverage from Clubic describes the launch as French AI daring to challenge GPT-6 Astra, while the open-weight release on October 27 is framed as the differentiator — something no US closed model offers. Mistral says it has narrowed the gap with the most capable systems; Quartz reports the company’s claim that ML4 leads open-weight models outside China by a substantial margin. The strategic pitch is simple: comparable frontier capability, but with published parameters. Whether the gap has truly narrowed is the open question.
What Are the Costs and API Availability?
Developers can already access Mistral Large 4 through a public API preview that launched on Tuesday, October 6. Third-party coverage of the model, including Kingy AI’s breakdown of specs and pricing, notes that preview API pricing and comparisons are being tracked alongside independent legal, finance, and coding evaluations. Mistral’s model page lists $1.36 and $0.68 per million input tokens, $0.14 and $0.07 for cached input, and $4.18 and $2.09 for output; see the page for current pricing context and conditions, and note that the open-weight release is planned for October 27, 2026. That gives teams an API preview now and an announced open-weight release later. Teams can use the preview to evaluate their workloads before making a production decision.
Why Was the Model Trained on European Infrastructure?
Mistral’s launch page says it trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.
Quartz reports 4,000 GPUs. That detail is central to the model’s European positioning. Training on company-owned infrastructure is part of Mistral’s pitch for European AI sovereignty.
What Do Independent Benchmarks Say So Far?
The independent picture is promising but incomplete. In the independent Intelligence Index, Large 4 makes what The Decoder describes as a big leap forward for Mistral. In the company’s own tests, Large 4 beats GLM 5.2; Quartz reports Mistral’s claim that ML4 leads open-weight models outside China by a substantial margin. However, multiple sources stress the same caveat: independent benchmarks of the full release are still missing. Mistral is also particularly effective, per the company, at tasks related to cyber, coding, manufacturing, finance, and multimodal work — all vendor-reported figures. The independent verification will come after October 27, when the weights are public and third parties can run their own evals.
Should Developers Start Building on the Preview Now?
For teams working in Mistral’s highlighted strength areas — cybersecurity, coding, manufacturing, finance, and multimodal applications — the public API preview is a low-risk way to evaluate a 1.05-trillion-parameter MoE model with 49 billion active parameters and a 1 million token context window listed by Mistral. The open-weight release is planned for October 27. Teams evaluating the API should verify separately whether the final model, license, and deployment options meet their requirements. There is one caution: since independent benchmarks are still missing, production decisions should wait for third-party verification of the vendor’s claims. Build prototypes and benchmarks now. Hold the production commitment until the community validates what Mistral promises.
Frequently Asked Questions
When will the Mistral Large 4 weights be released?
Mistral plans to publish the trained parameters by the end of October; MarketScreener gives October 27, 2026. Until then, the model is accessible through its public API preview.
How many parameters does Mistral Large 4 have?
Large 4 is a mixture-of-experts model with roughly 1.05 trillion total parameters, of which 49 billion are active per token. Mistral lists a 1 million token context window and describes the model as natively multimodal.
Why is the model called Le Chonk?
Mistral executives said the Le Chonk nickname deliberately nods to the community that had been rooting for the company to build a massive frontier model. It is an affectionate, self-aware joke rather than a formal product name.
Does Mistral Large 4 beat competing open models?
In Mistral’s own tests, Large 4 beats GLM 5.2. Quartz reports Mistral’s claim that it leads open-weight models outside China by a substantial margin. Independent benchmark results are still limited, so those claims need broader verification.
Summary
- Mistral Large 4 (Le Chonk) is a 1.05-trillion-parameter MoE model with 49 billion active parameters per token, a 1 million token context window, native multimodality, and open weights planned for release.
- Mistral says it trained Large 4 on 3,800 NVIDIA Grace Blackwell GPUs in Europe; Quartz reports 4,000. Quartz reports Mistral’s claim that Large 4 leads open-weight models outside China by a substantial margin.
- The public API preview launched October 6, 2026; MarketScreener reports the open-weight release for October 27.
- Mistral’s own tests show it beating GLM 5.2, while The Decoder reports a big leap in the independent Intelligence Index; broader independent benchmarks are still pending.
- Strength areas per the company: cyber, coding, manufacturing, finance, and multimodal tasks.
If you’re evaluating frontier models with an open-weight requirement, mark October 27 on your calendar and start testing the preview API today.