Private & On-Premise AI

Open-weight models are published with their weights, so you can run them yourself instead of calling someone else’s endpoint. Mistral, Llama, Qwen and GLM are among the families we work with, running inside your network under your access controls.

Two things bring people here. Legal, where the data cannot leave: patient records, privileged documents, anything covered by a contract you signed with your own customers. And cost, where per-token pricing that was fine at pilot volume stops being fine at production. When the constraint is legal, self-hosting isn’t a decision. The only question is which model fits the hardware you have.

What you get is a model chosen and tested against your task, deployed on your infrastructure, a configuration you own and can reproduce, and measured figures for speed and quality.

Model selection starts from the hardware, not the leaderboard. We size candidates to the memory actually available, pick the serving stack and quantisation format that suit the model and the workload, and discard anything that benchmarks well but won’t load.

Quantisation is a trade, and we measure it rather than assume it. Every step down in precision costs something, and where it costs you depends on the task. We test one variable at a time against your documents and your questions, then lock the result into a script you can rerun. On one document extraction build, vision-language models reached 83 tokens per second while handling layout, tables and handwriting.

That locked configuration is also how you avoid being frozen. Open models improve every few months, and because the benchmark is yours, a new release is tested against the same task and the same questions. You swap on evidence rather than guess.

We will tell you not to do this. Often enough self-hosting is the wrong answer, and a hosted API under a data processing agreement costs a fraction of dedicated infrastructure. Saying so loses us the larger project.

An open-weight model you can run is not the frontier model. On broad reasoning it will be behind. On a narrow, well-specified task it usually isn’t, and we say which you have before anyone buys hardware.

Where the output carries consequences, the system produces a draft. On federal regulatory filings for a US healthcare provider, every sentence traced back to something somebody had actually observed, and a reviewer edited and filed it. The model filed nothing.