Trim

Trim is an early-stage AI foundation model that simulates real-world physical systems in real time. Given initial conditions—like the starting position of waves on a beach—the model generates how those waves evolve over time. Instead of running a conventional solver at runtime, Trim trains on classical physics simulation data and learns to reproduce those outputs at inference time. The result is a model that acts like a constant-time lossy lookup table: it trades exactness for speed, delivering approximate physics predictions orders of magnitude faster than traditional methods. The core technical advantage is the Trim Transformer, a custom linear-attention architecture inspired by Galerkin-type attention. This design lets computation time scale linearly with both grid size and the number of dimensions, whereas traditional solvers scale polynomially in grid size and exponentially in dimensions. That means latency-critical tasks—such as an autonomous vehicle choosing its path—can drop by several orders of magnitude, and previously infeasible tasks like detecting gravitational waves become possible. As of mid-2026, Trim is still an early research project. There is no public API, no pricing, and no official documentation beyond blog posts that explain the architecture and design challenges. The project is open to researchers and engineers willing to engage at a research level. If you need guaranteed accuracy and support, traditional solvers or other surrogate modeling approaches are better choices.