Mistral launched Mistral Large 4, a roughly one-trillion-parameter model, as a public preview API on October 6, 2026, and says the open weights will follow at the end of October. Hugging Face lists October 31, and some reports say October 27. When the weights land, it will be one of the largest models anyone can download and run themselves.
The numbers
- About 1.05 trillion total parameters in a mixture-of-experts design, which means only a fraction of the model works on each request. Reports put the active count at around 49–52 billion, and the exact figure is still inconsistent across official pages.
- One-million-token context window, plus a vision encoder, so it takes images as well as text.
- Preview API pricing is reported at $1.36 per million input tokens and $4.18 per million output tokens, far cheaper than closed flagship models.
- Benchmarks so far are Mistral's own and preliminary. Wait for independent tests.
One outlet has disputed that the model exists at all, but that conflicts with Mistral's own announcement, so I'd give it little weight.
Open weights doesn't mean "free to run"
A trillion-parameter model needs serious GPU hardware even with mixture-of-experts. For most businesses, "open weights" will mean:
- Choice of host. You can run it on your own cloud account or with an inference provider, and you aren't locked into one vendor's API.
- Data control. For legal, healthcare or finance work, keeping data on infrastructure you control can matter more than raw quality.
- Fine-tuning. You can adapt the model to your domain. The licence terms will matter, though, and coverage notes they were still unclear at announcement.
When self-hosting makes sense
For a startup or small business: almost never at first. Use an API, measure your volume, and revisit when your monthly bill is large enough that dedicated GPUs pay for themselves, or when data rules force your hand. The useful part of open weights is that the option exists, which keeps API prices competitive.
How we approach it
At NextGen Fusion, our AI and software development company in Lucknow, we build AI features so the model behind them can change. Start on an affordable API, keep your own evaluation set, and move to an open-weight model like Mistral Large 4 when cost or privacy justifies it.
Planning a custom AI product? Here's how I build them.
