“MatX employs an ML research team that trains small LLMs from scratch to co-design and validate hardware-level decisions regarding numerics and attention mechanisms.”