Mistral Large 4 Preview: What Is Confirmed, What Is Promised, What Is Missing
Mistral opened a hosted preview of its trillion-parameter Large 4. The weights, license, and independent benchmarks are still to come.
Jump to 6 sections
Mistral opened a public preview of Large 4 on October 6, 2026. The model is a roughly 1 trillion-parameter multimodal mixture-of-experts system with a one-million-token context window. Open weights are promised by the end of October, but no license has been published, and every benchmark so far comes from Mistral itself.
Open-weight launches are easy to misread, because the word "open" arrives before the files do. Mistral has done something in between: a hosted preview now, downloadable weights later, and a license still marked coming soon. Model Drop tracks launches as they happen, and this brief separates what is confirmed from what is promised and what is simply unknown.
It is written for developers weighing whether to build on Large 4 now, wait for the weights, or skip it.
What Mistral Actually Released
According to Let's Data Science's report, Large 4 is in public preview through Mistral Studio, a hosted service. It is a multimodal mixture-of-experts model aimed at coding, agentic workflows, and multimodal tasks, and its documented context window is one million tokens. The API supports function calling, structured outputs, and document question answering.
Downloadable weights are planned for the end of October, after additional safety work. They are not out yet. The report adds that Mistral described the launch as staged, not a complete unrestricted release.
A Hugging Face listing, covered by AI Weekly, reads "open weights drop at the end of the month," and the article infers an October 31 target from a countdown on the page. That date is a plan, not a commitment. Treat it as a target that could slip.
The model accepts text and image input and produces text. Mistral highlights visual grounding and agentic work, with demos that zoom into engineering drawings and satellite images, according to the preview coverage.
How Big Is Mistral Large 4, Really?
Large 4 is reported at about 1 trillion total parameters, with 49 to 52 billion active per token, but the sources disagree on exact figures. A mixture-of-experts model holds many specialist subnetworks and uses only a few for each token, so total size and active size are two different numbers.
The Braindetox analysis of the preview compares Mistral's announcement and documentation. The announcement says 1 trillion total and 49 billion active. The documentation says 1.05 trillion total and 52 billion active, plus a 1.6 billion-parameter vision encoder. The article notes that neither source explains the gap.
The Hugging Face listing offers a likely explanation, according to AI Weekly. It gives 49 billion active parameters per token, or 52 billion including embedding and output layers. That would reconcile the active counts, though it does not explain the total.
The practical reading is simple. Memory needs follow total parameters, not active ones. At one byte per parameter, a 1.05 trillion-parameter model needs roughly a terabyte just for the weights. At half a byte per parameter, it still needs around half a terabyte. That is our own back-of-envelope estimate, not a figure Mistral has published, and Mistral has not stated memory requirements for self-hosting.
In other words, this is not a model you will run on one workstation. Our guide to local inference servers covers the smaller models that are.
What Does the Preview Cost?
The hosted preview is priced at $1.36 per million input tokens and $4.18 per million output tokens in Mistral's announcement. The documentation shows a second column with prices cut in half: $0.68 input and $2.09 output. Cached input is listed at $0.14, or $0.07 in the second column.
The Braindetox article does not say whether the lower column is temporary or whether the announcement prices apply only to the preview. Read both columns as provisional.
Here is what that means for one request. A task with a 50,000-token prompt and a 5,000-token reply costs about $0.089 at the announcement rates: $0.068 for input and $0.021 for output. At the lower column it costs about $0.044. Run that task 10,000 times and the difference is roughly $445.
The model is hosted in Mistral's own European datacenters, according to the same analysis. For teams with data-residency requirements, that may matter more than the price.
For a wider view of where this sits against other labs, see our look at four labs, three price bands, and one pattern from September.
What Do Mistral's Benchmarks Show?
Mistral's benchmarks show a strong agentic coding and security model that trails the top closed models on at least one blind comparison. Every number comes from Mistral, and independent reproduction is not possible until the weights are public.
The table below lists a selection from the Braindetox summary.
| Benchmark | Large 4 result | Note |
|---|---|---|
| DeepSWE v1.1 | 61.7% | Software engineering |
| SWE-Atlas-QnA | 59.4% | Repository understanding |
| Terminal-Bench 4 | 28.3% | Terminal workflows |
| AutomationBench | 59.9% | Business workflows |
| Surge AI blind coding eval | 3.74 of 5 | Second of five; Claude Opus 5 scored 4.22 |
| Cybench | 93% | 40 security competition exercises |
The blind coding result is the most informative. In that comparison, human raters did not know which model wrote which answer, and Large 4 placed second of five. That is a respectable showing, but it sits clearly behind the top model.
Mistral also reports strong cyber results, including 93% on Cybench and a top-five position on the Artificial Analysis Cyber Index. It says it leads open-weight models outside China on that index. These claims are narrow and vendor-reported, so wait for outside testing.
The older caution applies. A benchmark without its conditions says little. Our explainer on which AI model benchmarks still mean anything lists the questions to ask.
What Is Still Missing Before You Can Build on It?
Several things are missing, and each one affects a build decision. The license is the biggest. The documentation lists it as coming soon, and the announcement uses the words open-weight, open-source, and open interchangeably. Those are not the same thing, and the terms decide whether you can use the model commercially.
Other gaps, per the Braindetox summary:
- Architecture details, promised later.
- Output token limits and batch pricing.
- The exact weight release date and distribution location.
- Post-training methodology.
There is also a moving-target risk. The reinforcement learning run is described as still in progress, so the final weights may behave differently from the preview API. Anything you test now is a test of a preview.
The staged approach reflects a trend toward caution on capable models. Mistral says it is red-teaming with cybersecurity leaders, vetted partners, and state authorities, who get reduced moderation. For how labs approach that work, read our piece on how labs stress-test a model before launch.
Should You Build on Mistral Large 4 Now?
Build a prototype if you need European hosting and strong multimodal or agentic performance and can tolerate change. Wait if you need a confirmed license, a self-hosted deployment, or stable behavior. The weights and the license are the two events to watch.
A cautious path looks like this. Use the preview to build an evaluation set on your own tasks. Keep your prompts portable so you can switch providers. Do not sign a long commitment until the license is published and independent tests land.
At Model Drop, our editors will update this brief when the weights and license arrive. Your next step is to write down five real tasks and test them on the preview this week.
