Savante post-trains open-source models into domain specialists.

A general-purpose model is sufficient everywhere and exceptional nowhere in particular. For serious work, sufficient is not enough, and the paths to a better model are blocked: the frontier API is generic, uncontrollable, and expensive at scale, and building a specialist in-house needs post-training talent that is scarce and mostly inside the labs. Savante closes that gap. We take an open model and post-train it, using a customer's own data and evaluations, until it is genuinely expert in their domain and measurably beats frontier models on the work that matters to them. The differentiating work is evaluation curation: defining rigorously what "good" means for a problem and measuring it honestly, then training until the model clears that bar.
Working with PhysicsWallah, Savante's RL post-training took a 7B open model (Qwen2.5-Math-7B) from 54.7% to 90.2% on JEE Main April 2025 mathematics, ahead of o4-mini, Gemini 2.5 Flash, GPT-4.1, and GPT-4o. The model and paper are public.
Savante is built by researchers from Microsoft Research and JP Morgan, with published research at NeurIPS and CVPR.

Savante

Org chart

Teams

Offices

Savante

Org chart

Teams

Offices

Related companies