Mistral Large 4: a 1-trillion-parameter open-weight model, weights due late October
Mistral Large 4 is a mixture-of-experts AI model that French company Mistral released as a preview on October 6, 2026. It has about 1 trillion parameters in total but uses only about 49 billion for each token, and it reads both text and images. For now it is available only through Mistral's API; the downloadable weights are due at the end of October, with several outlets reporting October 27. On Mistral's own preliminary tests it scored 62% on an agentic coding benchmark, just ahead of China's GLM-5.3 (61%) and DeepSeek-V4-Pro (57%), and Mistral calls it the strongest open-weight model outside China. The license terms have not been published, and all scores are self-reported
The three lines
- Scale — ~1T total, ~49B active parameters per token; trained on ~4,000 Grace Blackwell GPUs over two months
- Scores — 62% on agentic coding vs GLM-5.3 61% and DeepSeek-V4-Pro 57% (Mistral's own numbers)
- Pending — weights at end of October, license unpublished; first model from the €3B September round
Key questions
- Mistral Large 4 benchmarks
- **Roughly level with, or slightly ahead of, the best Chinese open models on Mistral's own preliminary tests.** | Benchmark | Large 4 | Comparison | |---|---|---| | DeepSWE v1.1 (agentic coding) | 62% | GLM-5.3 61%, DeepSeek-V4-Pro 57% | | Finch (finance) | 67% | DeepSeek-V4-Pro 67% | | Harvey legal agent | 15% | Kimi K3 13%, GPT-6 Astra 5% | | DIOR-RSVG (visual grounding) | 73% | GPT-6 Astra 68% | All vendor-measured, taken before reinforcement learning finished.
- When will Mistral Large 4 weights be released
- **Mistral says 'end of October'; several outlets report October 27.** | Item | Status | |---|---| | API preview | From October 6, 2026 | | Weights | End of October | | License | Not published (Large 3 was Apache 2.0) | | Security edition | Separate, less restricted build for vetted developers, security firms, governments |
- Can I run Mistral Large 4 locally
- **Not on a normal PC.** | Precision | Approx. memory for weights | |---|---| | 16-bit | ~2 TB | | 8-bit | ~1 TB | | 4-bit | ~0.5 TB (outside estimate) | A mixture-of-experts model computes with a fraction of its parameters but must hold all of them in memory.
Europe now has a trillion-parameter model of its own, and Mistral says anyone will be able to download it. On October 6, 2026, the Paris-based company released a preview of Mistral Large 4, nicknamed "Le Chonk." Mistral calls it the most powerful open-weight AI system outside China. For most of 2026 the top of the open-weight rankings has belonged to Chinese labs such as DeepSeek, Zhipu (GLM) and Moonshot (Kimi), so the news is less about a single benchmark and more about who is back in the race.
1. What was released: 1 trillion parameters, 49 billion at work
Large 4 is a mixture-of-experts (MoE) model. It holds about 1 trillion parameters, but for each token only about 49 billion are used, which keeps running costs far below what its size suggests.
| Item | Detail |
|---|---|
| Release | October 6, 2026 (API preview) |
| Total parameters | ~1 trillion (some reports: 1.05T) |
| Active per token | ~49 billion |
| Inputs | Text and images (vision encoder ~1.6B parameters) |
| Context | 1 million tokens (per some reports) |
| Training | ~4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers, about two months, from scratch |
| Languages | 160+, including all official EU languages and Korean |
| Weights | End of October (reported as October 27) |
The weights are not out yet. Today only the API works, and Mistral says reinforcement learning is still under way. Before release it plans to publish more on training, safety testing and evaluations. It will also give developers, cybersecurity firms and government agencies a separate edition with fewer restrictions and extra cyber features, a tiered approach that echoes the access program Anthropic announced the same week.
2. Where it lands: neck and neck with China's best
| Benchmark (area) | Mistral Large 4 | Main comparisons |
|---|---|---|
| DeepSWE v1.1 (agentic coding) | 62% | GLM-5.3 61% · DeepSeek-V4-Pro 57% · Qwen 3.8 Max 51% · Reflection Beam 44% |
| Finch/FinWorkBench (finance) | 67% | DeepSeek-V4-Pro 67% · GLM-5.3 65% · DeepSeek-V4.1-Flash 57% |
| Harvey Legal Agent | 15% | Kimi K3 13% · DeepSeek-V4-Pro 8% · GPT-6 Astra 5% |
| DIOR-RSVG (visual grounding) | 73% | GPT-6 Astra 68% · Kimi K3 55% |
| Dense200 (visual grounding) | 42% | GPT-6 Astra 42% · Kimi K3 29% |
A one-point lead in coding and a tie in finance is catching up, not pulling away. Mistral concedes Large 4 still trails frontier closed models in areas such as coding. Its clearest edge is visual work — locating objects in images — which matches the uses it highlights: assessing storm damage for insurers, inspecting power lines, monitoring crops and turning technical drawings into CAD models.
The release came days after U.S. lab Reflection AI shipped its 501-billion-parameter open-weight Beam, and as DeepSeek V4.1 Flash narrowed the U.S.–China gap to about 3% on one leaderboard. Open-weight competition is now a three-way contest among the U.S., China and Europe.
3. Why Mistral leans on open weights
European banks, manufacturers and governments are wary of sending data to U.S. clouds. With the weights in hand, they can run a model on their own servers. Mistral also offers a regional API, including an EU region.
| Item | Detail |
|---|---|
| Funding | First model from the €3 billion round closed in September 2026 |
| New compute | Coming online through the first half of 2027 |
| Customers | 125+ large firms, including Airbus, ASML and HSBC |
| Co-training | Some customers helped train Large 4 via Mistral's Forge platform |
| Next | A family of specialized models built on Large 4 |
CEO Arthur Mensch, speaking in Abu Dhabi, said the narrative that Europe cannot compete "is something that is not true." Co-founder and chief scientist Guillaume Lample said the cyber-defense capabilities will let enterprises and governments defend themselves.
There is a practical catch. An outside estimate puts the weights at about half a terabyte even at 4-bit precision, far beyond a home PC or a small GPU server. In practice this is an open-weight model for organizations that own data centers.
4. What remains unconfirmed
- License: Mistral's docs list Large 4 in the "Open" category (version 26.10) but give no terms. Large 3 was Apache 2.0; a custom license would narrow what "open" means.
- Date: Mistral says "end of October"; October 27 is a press report.
- Scores: Self-measured mid-training. Independent results will only come after the weights are out.
- Specs: Reports differ on total parameters (1T vs 1.05T) and training GPUs (about 3,800 vs about 4,000).
- Claims of beating Anthropic's Fable 5.1 on vision and open models from Kimi, DeepSeek and Meta on cyber tasks came without numbers.
Sources
- The Next Web — Europe's Mistral launches Large 4 to challenge China's lead in open AI models
- MarkTechPost — Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE
- MarketScreener — Mistral unveils Large 4, its new open-weight AI model
- Technology.org — Mistral Large 4 Arrives With 1 Trillion Parameters