Mistral Large 4 is a trillion-parameter model you cannot download yet
Mistral opened a preview of its 1T-parameter mixture-of-experts model and promised the weights by the end of October. Here is what the numbers say so far.
- Model
- Mistral Large 4
- Maker
- Mistral AI
- Size
- 1.05T total, 52B active
- Context
- 1M listed, 524K served
- License
- Not announced
- Weights
- Promised
Mistral put Mistral Large 4 into public preview on October 6. You can call it through the API on Mistral Studio today. The part that makes it a story for this site, the open weights, is promised “by the end of the month” in Mistral’s announcement, with no license named yet.
So for now this is a hosted preview with a promise attached.
How it works
Large 4 is a mixture-of-experts model. Instead of pushing every token through every weight, a router picks a small group of specialist sub-networks for each token and leaves the rest idle.
Mistral’s announcement describes 1 trillion parameters with 52 billion active. Its model card puts the total at 1.05 trillion, plus a 1.6 billion parameter vision encoder, according to Developers Digest. Launch-day coverage said 49 billion active; Mistral’s current figure is 52.
That split matters in two different ways:
- Speed and cost track the active count. Each token costs roughly what a 52B model would.
- Memory tracks the total. You still have to hold all 1.05 trillion parameters somewhere.
Developers Digest estimates the weights alone at about 1 TB at 8 bits per parameter and about 525 GB at 4 bits. That is a multi-GPU server, not a workstation. Open weights here means “your company can host it”, not “it runs on your desk”.
Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters, and that reinforcement learning is still running. The scores below may move before release.
The benchmarks
These are Mistral’s own numbers from the announcement.
| Benchmark | Mistral Large 4 | Comparison given |
|---|---|---|
| DeepSWE v1.1 | 61.7% | None |
| SWE-Atlas-QnA | 59.4% | None |
| Terminal-Bench 4 | 28.3% | None |
| Coding Agent Index | 49.8% | None |
| AutomationBench | 59.9% | None |
| Surge AI blind coding eval (1 to 5) | 3.74 | Claude Opus 5: 4.22, GLM-5.3: 3.60, Kimi K3: 3.59, GLM-5.2: 3.40 |
| Dense 200 visual grounding | 42% | GPT-6 Astra: 41% |
| Cybench | 93% | None |
Most rows come without a competitor’s score beside them, which makes them hard to place. The fullest head-to-head, a blind human coding evaluation, has Large 4 second out of five: behind Claude Opus 5 and ahead of the three open models it was compared with.
The independent picture is thinner. Artificial Analysis gives the preview 38 on its Intelligence Index, against a median of 26 for models at its price, as reported by Developers Digest. The same write-up notes that Mistral Large 3 scored 9, so the jump between generations is large even if 38 is not a chart-topping number.
Loose ends
- Context window. Mistral’s card lists 1M tokens. Artificial Analysis and at least one gateway list 524K for the preview as served. Plan around the smaller number until that is settled.
- License. Not announced. Whether you can use the weights commercially is the first thing to check on release day.
- Price. List price is $1.36 per million input tokens and $4.18 per million output tokens, currently discounted by half during the preview.
Try it yourself
The preview is on Mistral Studio. There is nothing to download yet, and we have not run the model ourselves: this is based on Mistral’s announcement and third-party reporting. We will update the Open Weights tracker when the files exist.