Trim

Trim develops the Trim Transformer, a linear-attention architecture that trains generative AI models to simulate physical systems evolving over time. It uses multi-linear attention—computing attention as QK^TV with O(nd²) complexity—to avoid the quadratic scaling of standard softmax attention on long sequences and large grids. The company trains these models by running traditional physics simulations and feeding the results into its pipeline, aiming to reduce the compute cost and latency that grow with simulation dimensions and grid size.

The technology targets researchers and engineers working on high-dimensional physics simulations, including climate modeling, semiconductor and battery design, fusion plasma modeling, molecular bonding research, gravitational-wave detection, and autonomous-vehicle pathing. Trim distributes its work as an open-source PyTorch-compatible Python package installable via pip and competes with conventional numerical simulation methods and standard transformer implementations. It positions on linear scaling and lower memory usage.