Hamiltonian Research is an independent AI lab building attention-free language models using state-space dynamics (Mamba-2) and latent diffusion instead of transformers. The flagship DIMBA architecture claims constant-memory O(1) inference and 44x CPU speedups, currently training a 135M parameter model with results pending.

One insight: This is deep foundational AI research (novel architecture, published paper, physics-grounded claims), not a product wedge: replicating the core innovation requires months of ML research and novel training infrastructure, not fast AI-assisted coding, making it strategically irrelevant for a studio seeking buildable 1-2 week launches.