Mistral Large 4
A roughly trillion-parameter multimodal model released open-weight — the gap between hosted and self-hosted narrows again.
Mistral’s trillion-parameter multimodal model, released with open weights. Like Beam it is a mixture-of-experts, so the compute per token is a fraction of the total size. Multimodal means it takes images alongside text, not just prose.
What you build with it
- Running a frontier-class model entirely inside your own infrastructure
- Regulated work where data is not allowed to reach a third-party API
- High-volume inference where API cost has overtaken the cost of hardware
- European data-residency requirements
What it gets you
- Open weights at genuinely frontier scale
- Natively multimodal rather than text-only
- ~52B active parameters of 1.05T total, so it is servable
- Parameters
- 1.05T total · ~52B active
- Architecture
- Multimodal MoE
- Codename
- Le Chonk
- Weights
- Open — announced for later in October
Straight from the source
Watch instead of reading
The fast version, if you do not have time for the whole write-up.
Why it matters
The argument for hosted frontier models has always been that nothing you can run yourself is close enough to matter. Every open-weight release at this scale chips at that, and it changes the shape of the question for anyone with data that cannot leave their own infrastructure.
Open weights at this scale are a procurement fact, not a hobby project
A trillion-parameter model is not something you run on a workstation. What open weights buy at this size is the right to run it inside your own boundary — your cloud account, your region, your audit trail, your data never crossing to a third party.
For regulated work that is often the only thing standing between a project and approval. The model does not have to beat the best hosted model; it has to be good enough while being somewhere the data is allowed to go.
It is also a mixture-of-experts: 1.05T parameters in total but only around 52B active per token. Same lesson as Beam — the total tells you the memory bill, the active count tells you the compute bill, and those are very different numbers.
What "open weights" does and does not mean
Open weights means you can download and run the parameters. It does not automatically mean the training data is published, the licence permits every commercial use, or that you may fine-tune and redistribute freely.
Read the actual licence before building on it. The range across recent "open" releases is wide — genuine Apache 2.0 or MIT at one end, and restricted community licences with revenue thresholds and use limits at the other.
Verify this one before you plan around it
At the time this entry was written the weights were reported as promised by month-end rather than already published, and the parameter count came from press coverage rather than a model card.
Treat the specs above as a signal to watch, not a committed fact. Check Mistral's own release page and the Hugging Face model card before anything depends on it.
Reach for it when
- Data residency or privacy rules out sending data to a hosted API
- You have, or can rent, serious GPU capacity and the ops skill to run it
- Hosted per-token costs at your volume have passed the cost of running your own
Watch out for
- Weights were reported as pending, not shipped — confirm before planning
- Read the licence; "open weights" covers a wide range of permissions
- Serving a model this size well is its own engineering project, not a deployment step