Independent AI research lab · Abu Dhabi

Rebuilding the language model from its equations of motion.

Attention scales quadratically; physics doesn't. We build language models on state-space dynamics and continuous latent diffusion. Constant-memory inference, with reasoning that refines before it speaks.

Read the paper Our research

Research program

Three commitments, one architecture.

The transformer's attention mechanism buys expressivity with quadratic cost. Our program replaces it end to end: not by approximating attention, but by changing the underlying dynamics of the model.

State-space backbones

O(n) time · O(1) state

Selective state-space models (Mamba-family) carry context as an evolving fixed-size state rather than a growing cache. Sequence length stops being the enemy: linear-time processing, constant memory, effectively unbounded context.

Continuous latent diffusion

refine, then emit

Instead of committing to one token at a time, the model refines a continuous latent representation of its answer through diffusion steps. Iterative reasoning happens in latent space, before anything is projected into text.

Constant-memory inference

no KV-cache

Eliminating the KV-cache collapses inference into a fixed memory footprint. Generation speed improves by an order of magnitude, making real-time complex reasoning practical on modest hardware.

Flagship architecture

DIMBA

DIMBA fuses a bidirectional Mamba-2 state-space denoiser with cosine-scheduled Gaussian diffusion over latent text states: a non-autoregressive model that generates whole sequences in one pass instead of token by token. The library ships classifier-free guidance, self-conditioning, cross-architecture distillation from any Hugging Face transformer, and DPO/SimPO/GRPO post-training. It runs without CUDA, on CPU and Apple Silicon included.

A 135M-parameter training run is underway. First results are coming soon.

Paper · DOI 10.55277 Model · Hugging Face

backbonebidirectional Mamba-2 SSM

decodinglatent diffusion · one pass

conditioningAdaLN-Zero + CFG

scale135M in training · scaling up

backendsCUDA · MPS · MLX · CPU

inference speed44x CPU (MLX, measured)

kv-cachenone · O(1) memory

status135M run in progress

The lab

Founded by Faris Allafi.

Hamiltonian Research is an independent lab based in Abu Dhabi, founded and led by Faris Allafi, architect of the DIMBA model and, at 13, one of the youngest published researchers working on post-transformer architectures.

The lab's conviction is simple: the current generation of language models is defined by an architectural accident, not a law of nature. Better dynamics exist, and they are worth building from first principles: openly, with technical depth as the only currency.

Formerly operating as DimbaLabs.

Faris Allafi, founder of Hamiltonian Research