A busy spring
The first months of 2026 brought a string of open-weight model releases: Gemma 4 from Google, published under the Apache 2.0 licence in sizes ranging from edge devices to more than thirty billion parameters; Llama 4 from Meta; new generations of Qwen and Mistral. Each of these families now offers versions that run comfortably on a server with one or two GPUs.
Two years ago, running a “serious” model locally was an experiment. That is no longer the case.
The right question is not “the biggest model”
Public leaderboards compare models on general tests: competition mathematics, encyclopaedic knowledge, puzzles. That is not what your teams do. They look for a clause in a contract, summarise minutes, write a letter, extract the fields of an invoice.
That is why we run our own tests: more than 360 tasks built on real professional work, run across more than 25 models. On these tasks, the gap between the best cloud model and the model we deploy on a local server is a few points, and it does not show in day-to-day use. The details are on the MIA page.
What open-weight models make possible
- Sovereignty: the model runs on your premises, no data leaves your network.
- A cost that no longer depends on usage: no per-token billing, adoption does not drive up the bill.
- Specialisation: an open model can be fine-tuned on your vocabulary and use cases.
- Independence: you can switch models when a better one comes out, without changing tools.
Licences: read the terms
“Open” does not always mean the same thing. Some models are released under a free licence such as Apache 2.0, which allows commercial use without particular restrictions. Others come under their vendor’s own licence, which may limit certain uses or impose conditions above a given number of users. Before deploying a model in production, its licence deserves the same attention as its performance.
Hardware is no longer an obstacle
Progress does not only come from the models. Mixture-of-experts architectures, which activate only part of the parameters for each request, and quantisation techniques, which reduce the memory required, make it possible to serve capable models on a single server. Inference engines have improved too, with throughput that can serve several dozen simultaneous users.
For a company, this means local AI requires neither a data centre nor a research team: one properly sized server, installed on your premises or with your host, is enough to cover the uses of an entire organisation.
The right model, in the right place
The real work is therefore no longer to find a powerful enough model, but to choose the right size for each use: a compact, fast model for document search, a specialised model for code, a larger model for agent tasks. That is what we call frugal AI: just the right power, in the right place, at a controlled cost.
Our local API exposes these models behind an interface compatible with OpenAI and Anthropic: your existing tools plug in unchanged, and the model can evolve without your teams having to worry about it.
