Applied artificial intelligence, neural systems, and smart agents.
Fusing generative AI, multi-agent frameworks, and high-throughput data processing to automate business workflows.
Mistral Large 3
Large 3 is the heaviest unit whose weights are published in this reference: 675 billion total parameters and 41 billion active parameters, a 256k-token context window, under Apache 2.0. This configuration enables deployment on owned infrastructure and commercial operation without a separate licence.
- 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. Change log
Two parameter figures, two distinct architectural constraints
The specification for this unit reports two parameter figures, and typically only one drives the decision: 675 billion total, 41 billion active. These two figures constrain two different components of system load, and conflating them is costly.
41 billion active parameters set the compute load required to produce each token, roughly equivalent to a 41-billion unit. 675 billion total parameters set the memory requirement: all of those weights must be loaded and addressable so the correct segment can be invoked at each computational step. So this unit behaves like a mid-tier model on throughput and like a very large one on memory footprint. Hardware provisioning based on the 41-billion figure alone, expecting deployment on an ordinary graphics card, underestimates the required capacity.
The licence: the most consequential specification of this unit architecture
Apache 2.0 is a standard, industry-recognised licence: unrestricted commercial operation, no user ceiling, no mandatory attribution display, no naming condition on derivatives. Set this licensing configuration against Llama 4, which imposes a 700-million monthly-active-user ceiling and dictates the naming of any derived model, or against Mistral own Voxtral TTS, whose licence text forbids commercial operation outright.
For an organisation deploying this unit inside its own infrastructure, that licensing distinction carries more decision weight than any benchmark figure.
A 256k context window: mid-range in this dataset
256k is not a weak figure, but it does not sit at the high end of this reference. Opus 5, Fable 5, Sonnet 5 and DeepSeek V4 are all provisioned at one million tokens, and Llama 4 Scout reaches ten million. For a workload that requires processing one very long document in a single pass, this unit is not the first choice, and that assessment is stated plainly.
A data gap: pricing
Mistral publishes no per-million-token rate for any of its models and instead sells subscription tiers, so the price column for this unit remains empty in our listing. Under a self-hosted deployment this parameter has no bearing at all, since the operating cost equals the hardware cost — which may itself be why Mistral does not publish a rate.
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: The network path to the Mistral API has not been measured from Iran, and no citable supported-countries list was identified; accordingly no data is recorded on the hosted service layer for this unit. Its weights, however, are published under Apache 2.0, and a model with open weights deploys on owned infrastructure, independent of any account or country list. This finding is derived from the licence text, not from a network measurement.
Documented strengths
- Apache 2.0 with no user ceiling and no naming condition, a rare configuration at this scale
- Published weights, enabling deployment on owned infrastructure with no dependency on a country list
- 41 billion active parameters, so the compute cost per token runs far below what the 675-billion total figure suggests
Room for improvement and technical considerations
- 675 billion parameters must fit in memory in full; the 41-billion active figure does not reduce this requirement, and the hardware need is substantial.
- The 256k context window is mid-range in this reference, against one million on several other units and ten million on Llama 4 Scout.
- Mistral publishes no per-million-token rate, so an API workload cannot be costed in advance.
- Persian is not named among the languages the model specification lists. That list is explicitly non-exhaustive, so the conclusion recorded is not unsupported; it is unnamed.
Technical verdict
For deploying a frontier-class unit inside owned infrastructure where the operating licence is the deciding factor, this is our first-choice configuration. For the top coding score or the largest context window, evaluate other options.
Frequently raised questions, with documented answers
With 41 billion active parameters, does this unit run on a single graphics card?
No. 41 billion is the per-token compute figure, not the memory figure. Running it requires all 675 billion parameters loaded, since any segment may be invoked at any computational step.
Does Apache 2.0 genuinely permit commercial operation?
Yes. Apache 2.0 is a standard licence with no user ceiling and no naming condition. Note this licence is not uniform across every Mistral unit: Voxtral TTS forbids commercial operation.