The Mistral Large 4 open weights are now public, which means you can download a 1-trillion-parameter model and run it on your own hardware. The catch is a modified MIT license: you can build products with it, but selling it as a hosted service needs a separate deal with Mistral.
Here’s what changed, what it costs through the API, and who should actually bother self-hosting it.
Quick facts first:
- Size: 1 trillion total parameters, mixture-of-experts, with roughly 50 billion active per token.
- License: Mistral Ecosystem Open-Weight license (modified MIT).
- Weights: on Hugging Face as
Mistral-Large-4-1T-A52B. - API price: $0.70 per million input tokens, $2.10 per million output.
- Timing: public preview in October, weights released around late October.
Those five lines cover most of what you need. The rest of this post explains the fine print.
What Mistral actually released
Large 4 is a natively multimodal model, so it handles images as well as text. Mistral first put it into public preview in October, then opened the weights after.
The name on Hugging Face ends in “1T-A52B,” which points to about 52 billion active parameters. Early coverage from Let’s Data Science cited roughly 49 billion. Mistral’s own page doesn’t state the number, so treat “around 50 billion” as the safe summary.
That mixture-of-experts design is why a 1T model is usable at all. Only a slice of the network runs for each token, so serving cost tracks the active size more than the headline number. Memory is another story, since every expert still has to be loaded.
Where Mistral Large 4 stands on benchmarks
These numbers come from Mistral’s announcement, so they are self-reported and worth independent testing before you commit.
| Benchmark | Reported result |
|---|---|
| Cognition FrontierCode 2.0 | 19.7% (top of Mistral’s chart) |
| Artificial Analysis Coding Agent Index | 48%, tied with GLM-5.3, using about 47% fewer tokens |
| Terminal-Bench 4.0 | 41%, up 13 points from the preview |
| BrowseComp | 84.5% |
| FinWorkBench | 73.3% |
The token-efficiency claim is the interesting one. Matching a rival while using nearly half the tokens means lower bills on long agent runs, if it holds in your own workloads.
What the license lets you do
Mistral calls it a modified MIT license. Use, modification, fine-tuning, and commercial use are all allowed.
The restriction is narrow but real. If you offer the model to others as a Model-as-a-Service, you need a separate commercial agreement. Let’s Data Science flagged this as the point resellers should read twice.
For a company using Large 4 inside its own app or internal tools, that clause likely never applies. For an inference startup, it’s the whole question.
Self-hosting vs. the Mistral Large 4 API
Here’s how the two routes compare for most teams.
| Self-host the weights | Use the API | |
|---|---|---|
| Setup | vLLM or SGLang; TensorRT-LLM for high throughput | Mistral API, Studio, OpenRouter, NVIDIA, Microsoft Foundry |
| Cost | Your own GPUs; Mistral lists no hardware spec | $0.70 in / $2.10 out per million tokens |
| Data control | Stays on your servers | Goes to the provider |
| Safety layers | Yours to build | Hosted guardrails included |
That last row matters. As noted by Let’s Data Science, hosted guardrails don’t ship with a self-hosted copy, so you’ll need your own safeguards. If you’ve worked through vLLM on Kubernetes, you already know how much plumbing that involves.
Should you run it yourself?
Probably not on a laptop. A 1-trillion-parameter model needs a multi-GPU server, and Mistral hasn’t published exact hardware needs. If you want something small enough for a desktop, our guide to running Qwen3 locally is a better starting point.
Self-hosting makes sense when data can’t leave your network, or when your volume is high enough that per-token pricing hurts. Everyone else can try the API first and keep the weights as an exit option.
How to test it before you commit
Don’t pick a model from a chart. Take ten real tasks from your own work, such as a bug fix, a long document summary, or a data extraction job, and run them through the API first.
Track three things: whether the answer was right, how many tokens it used, and how long it took. The token count is where Mistral’s efficiency claim will show up or fall apart.
If the API results look good and your monthly bill is climbing, that’s the moment to price out a GPU server. Doing it in that order keeps you from buying hardware for a model that turns out not to fit.
Frequently Asked Questions
Is Mistral Large 4 really open source?
It’s open-weight, not classic open source. The license is a modified MIT with a restriction on offering the model as a hosted service.
Can I use Mistral Large 4 commercially?
Yes, for your own products and tools. Reselling it as a Model-as-a-Service requires a separate agreement with Mistral.
Where can I download the weights?
On Hugging Face under mistralai/Mistral-Large-4-1T-A52B, according to Mistral’s announcement.
How much does the Mistral Large 4 API cost?
$0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens.
My take: this is a strong release for teams with real GPU budgets and a reason to keep data in-house. For everyone else, the cheap API price makes it a low-risk model to test, and the open weights are a handy safety net if pricing or terms ever change. Run your own tests before trusting any chart, mine included.
Leave a Reply