Mistral Large 3
Large 3 is the biggest model in this reference whose weights are published: 675 billion total parameters, 41 billion active, a 256k context window, under Apache 2.0. You can take it, host it yourself and use it commercially without asking anyone.
- Apache 2.0
- 675 billion parameters
- 256k context window
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Two parameter counts, two different limits
The model card gives two numbers and people usually read one. 675 billion total, 41 billion active. Those two numbers constrain different things and confusing them is expensive.
41 billion active means the compute needed to produce each token is roughly that of a 41 billion parameter model. 675 billion total means all of those weights have to be loaded somewhere, so the right part can be reached at each step. So this model behaves like a mid-sized one on speed and like a very large one on memory. Anyone who sees 41 and assumes it will come up on an ordinary graphics card orders the wrong hardware.
The licence, which is this model most important spec
Apache 2.0 is a standard, well understood licence: commercial use is free, there is no user ceiling, no mandatory attribution display, no naming condition. Compare that with the Llama 4 licence, which caps you at 700 million monthly active users and dictates the name of any model you derive, or with Mistral own Voxtral TTS, which forbids commercial use outright.
For a company that wants to bring a model inside its own infrastructure, that difference changes the decision more than any benchmark number.
A 256k window, which is mid-table here
256k is not a bad number but it is not a high one in this reference. Opus 5, Fable 5, Sonnet 5 and DeepSeek V4 are all at a million, and Llama 4 Scout is at ten million. If your job is reading one very long document in a single pass, this is not the first choice, and that is the honest answer.
What we cannot tell you
The price. Mistral publishes no per-million-token rate for any model and sells subscription tiers instead, so the price column for this model is empty in our list. If you host it yourself the question does not arise at all and your cost is hardware, which may be Mistral own reason for not printing it.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- not checked
- Payment
- not checked
- Free tier
- no
How we checked: We have not tested the Mistral API from Iran and found no supported-countries list we could cite, so we say nothing about the hosted service. But the weights of this model are published under Apache 2.0, and a model with open weights runs on your own hardware with no account and no country list. We read that in the licence, not in a network test.
What it is good at
- Apache 2.0 with no user ceiling and no naming condition, which is rare at this size
- Published weights, so it runs inside your own infrastructure and depends on no country list
- 41 billion active parameters, so the compute cost per token is far below what 675 billion suggests
Where it falls short
- 675 billion parameters have to fit in memory. The 41 billion active figure does not change that, and the hardware requirement is serious.
- The 256k window is mid-table here, against a million on several other models and ten million on Llama 4 Scout.
- Mistral publishes no per-million-token price, so an API workload cannot be costed in advance.
- Persian is not among the languages the model card names. The list is explicitly not exhaustive, so we do not say it is unsupported; we say it is not named.
Our take
If you want a frontier-class model inside your own infrastructure and the licence matters to you, this is our first choice today. If you want the top coding score or the largest context window, look elsewhere.
Questions people actually ask
With 41 billion active parameters, will it run on one graphics card
No. 41 billion is the compute figure per token, not the memory figure. To run it, all 675 billion parameters must be loaded, because any part of them may be reached at any step.
Does Apache 2.0 really mean I can use it commercially
Yes. Apache 2.0 is a standard licence with no user ceiling and no naming condition. Note that not every Mistral model carries it: Voxtral TTS forbids commercial use.