01
Foundation models that learn physics from physics
We are building scientific foundation models trained to reason about simulation and to write the language our solvers execute.
Why
A general language model has read about physics; it has not run any. Simulation generates unlimited correctly-labelled data on demand, and every answer can be checked against a solver — so the reward signal is verifiable rather than approximated.
Approach
FPS-SciFM, with variants on Granite 4.1-8B-instruct and other architectures, trained by reinforcement learning from verifiable rewards (RLVR). FLUX is tokenizer-optimised and our tools curate training files automatically, so the corpus grows from our own simulation output.
State
In continuing development. FPS-SciFM v1.0 model launched on SimuPort.
