Mistral released a 1-trillion-parameter open-weight model this morning, October 6, 2026, and called it "Le Chonk." The model - Mistral Large 4, abbreviated ML4 - is a 1.05-trillion-parameter open-weight multimodal mixture-of-experts, and it went live on the Mistral API as a public preview. The "open-weight" label is the thing teams are reacting to. But the actual question is whether you can do anything with that label today, and the answer is: not quite yet.
What "open weight" means for Mistral Large 4 right now
Open weight means the model files will eventually be downloadable and self-hostable, with no API call required. But "open-weight" does not yet mean "downloadable" - the actual model files developers would need to self-host Large 4 are not shipping on day one.
Mistral's Hugging Face upcoming-release page currently lists October 31 as the expected date , while VentureBeat and Euronews were given a specific date of October 27.
Until then, the only access is through Mistral's own API, which makes this a closed-access preview of a future open-weight model. That is a meaningful difference for any team whose reason to care about open weights is running inference on their own hardware.
The reinforcement learning phase - in which the model improves through trial and error and feedback - is still ongoing, and Mistral reports that gains have not begun to slow. The published benchmark results are preliminary, and the company expects them to change before the weights are released. You are evaluating a moving target.
The number that actually matters: 49 billion active parameters
The headline is 1 trillion parameters. The number that determines your inference cost is different. Mistral Large 4 activates only 49 billion of its 1 trillion total parameters for any given task, indicating a mixture-of-experts style architecture. That is the same active count as DeepSeek V4 Pro - a model with 1.6 trillion total parameters.
This matters in two directions at once. On the one hand, MoE sparsity makes inference faster and cheaper per token than a dense model of equivalent size would be. At the standard tier, Mistral is pricing the API at $0.68 input and $2.09 output per million tokens. On the other hand, self-hosting a sparse MoE still requires loading all 1.05 trillion parameters into GPU memory - you pay the memory bill even for the experts you don't use on any given forward pass. A bf16 checkpoint of roughly 1.05 trillion parameters runs to around two terabytes of VRAM, putting full self-hosting in the range of a large H100 or Blackwell cluster rather than a team's on-premise server rack.
The model was trained from scratch over two months using about 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers, consuming roughly 10 megawatts of power during training. The sovereign data-center angle is deliberate: the launch is Mistral's clearest statement yet of its sovereign AI pitch. In a statement reported by AFP, Mistral said enterprises "across financial services, manufacturing and the public sector can govern the intelligence, data, compute and operations" without sending sensitive knowledge outside. Mistral calls it the first milestone funded by the €3 billion Series D that valued Mistral at more than €21 billion in September.
What the benchmarks say - and where to be skeptical
On agentic coding, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index of 49.8%, placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
The cybersecurity number is where Large 4 stands furthest apart from the field. Mistral reports 93% on Cybench and 82% on CyberGym-E2E, placing ML4 in the global top 5 on the Artificial Analysis Cyber Index. The more interesting claim is structural: Mistral states several frontier closed models score near zero on CyberGym-E2E because they refuse outright. For security teams evaluating the model for offensive-capability testing, this is the real differentiator - a capable model that won't refuse the task.
On broad capability, the picture is more measured. On Vals AI's broad industry index, it scores 48.05%, behind GPT-6 Astra, GPT-6.1 Sol, Claude Sonnet 5.5, and Claude Opus 5.5.
Comparing the two models most directly on neutral infrastructure: on the Artificial Analysis Intelligence Index v4.3, Mistral Large 4 Preview scores 38.4 versus DeepSeek V4 Pro 0813 at 36.0, with a cost per task of $1.13 vs $0.67. Mistral Large 4 wins the agentic-coding row by nearly thirteen points and the workflow-automation row narrowly, while DeepSeek V4 Pro wins open-ended knowledge by six points and costs 40% less per task.
That pricing gap matters if you are running high-volume workloads. Paying 68% more per task to win on agentic coding is a trade-off worth doing the math on before committing.
Beagle in action#engineering, evaluating models for an internal code-review agentWhat to actually do before October 27
| If your use case is… | Action now | Action after weights land |
|---|---|---|
| Agentic coding agents | Test the API preview on your real repo tasks | Compare self-hosted latency against the API cost |
| Legal or financial document work | Run bounded tests on representative docs | Check final license terms before building on it |
| Cybersecurity / red-teaming | Request early API access; this is the clearest strength | Evaluate whether your threat model needs on-prem weights |
| General assistant / chat | Probably not the right model; closed-model costs are competitive | Wait for independent benchmarks after the RL run completes |
The current public-preview phase gives developers API access while Mistral continues safety work, which limits what can be concluded about long-term availability, final safeguards, and independent performance. Developers evaluating the model should begin with bounded tests that use representative documents, tools, and workflow prompts rather than moving sensitive production tasks immediately.
The one thing not to do: build your evaluation around today's benchmark numbers. Mistral says the weights will ship by the end of October, and told reporters October 27. The release should come with a final checkpoint, because the reinforcement learning run behind the preview is still in progress. The model being benchmarked today is not the model that will land in your registry.
A tool like Beagle can draft the internal comparison note for your team as you test - pulling the benchmark tables from Artificial Analysis and the Vals Index into a single channel post for a quick review - but the evaluation itself is yours to run.
Mistral Large 4 open weight: common questions
When will Mistral Large 4 weights be released?
Mistral says by the end of October 2026. The company told reporters October 27; Hugging Face lists October 31. The API preview is available now. Expect the final release to come with a full architecture disclosure, additional benchmark results, and the named license terms, none of which have been published as of today.
Is Mistral Large 4 actually open source?
No, and it may not be. Mistral is using the label "open weight," not "open source." Open weight means the model files will be downloadable; it says nothing about the license permitting commercial use, modification, or redistribution without restrictions. The license terms have not been named. Treat it as open weight until the actual license lands alongside the weights.
How many GPUs does it take to self-host Mistral Large 4?
The 1.05-trillion-parameter sparse MoE checkpoint requires loading the full parameter count into GPU memory even though only 49 billion are active per forward pass. In bf16, that is roughly 2 terabytes of VRAM - around 16 to 20 H100 80GB GPUs at minimum, or a smaller count of Blackwell-tier cards. This is not a model for a single-server deployment.
How does Mistral Large 4 compare to DeepSeek V4 Pro on agentic tasks?
On independent Artificial Analysis benchmarks, Large 4 scores 38.4 on the Intelligence Index versus DeepSeek V4 Pro's 36.0, and beats it by about 13 points on Terminal-Bench 4 agentic coding. DeepSeek V4 Pro costs roughly 40% less per task. Both have 1 million-token context windows. Large 4 adds native image input; DeepSeek V4 Pro does not.
Should teams use Mistral Large 4 now or wait?
Use the API preview now only if you have a specific workload - legal, financial, cybersecurity, or agentic coding - and you can run bounded tests against your own representative tasks. Do not migrate production workloads or build internal infrastructure assumptions around a model whose training run is not finished and whose license is unpublished. The useful evaluation window is October 27 onward, when the final checkpoint, architecture details, and license terms are all on the table at once.