HAMILTONIAN / RESEARCHRethink the sequence.
built in Abu Dhabi
State-space dynamics meet masked diffusion. Constant-memory language models that generate by refining a whole answer—not predicting the next token.
A different machine for language.
DIMBA combines masked diffusion with a bidirectional Mamba-2 backbone. It sees the whole damaged sequence, reasons from both directions, then commits the most confident tokens and repeats.
Mamba-2
Linear-time sequence processing without an ever-growing KV cache.
Diffusion
Iterative refinement across the complete response in parallel.
The dial
Trade compute for quality continuously instead of choosing a fixed model.
Honest numbers beat perfect demos.
DIMBA’s first run lost the factual-knowledge test, but showed two structural strengths: reconstructing missing text and resisting repetitive loops.
| Model | Factual QA | Loop rate ↓ | Infill recovery ↑ |
|---|---|---|---|
| hr-diffuse-1-nano | 15.0% | 7.5% | 14.0% |
| SmolLM-135M teacher | 82.5% | 37.5% | 2.9% |
| SmolLM-135M-Instruct | 60.0% | 2.5% | 0.0% |
| GPT-2 | 20.0% | 90.0% | 0.0% |
| Pythia-160M | 10.0% | 15.0% | 1.7% |
40 factual questions · results reported in the July 2026 technical report · lower loop rate is better
Founded by Faris Allafi.
Hamiltonian Research is an independent lab based in Abu Dhabi, founded and led by Faris Allafi, architect of DIMBA and one of the youngest published researchers working on post-transformer architectures.
The current generation of language models is defined by an architectural choice, not a law of nature. Better dynamics are worth building openly, with technical depth as the currency.
Formerly operating as DimbaLabs.
