Research
What we are working on.
Six threads, from expert pruning through to population genomics. Each one has its own page with the detail, the numbers, and the published work it sits alongside.
- 01 · CompressionExpert pruningCutting experts out of large mixture-of-experts models until they fit the hardware people actually own.Read the thread Shipped · open weights
- 02 · SpeechMultilingual speech recognitionDialect-aware recognition for languages where labelled audio is scarce and the incumbent systems are far larger than they need to be.Read the thread Shipped · leaderboard-tier
- 03 · EvaluationRefusal ablationRemoving refusal behaviour from open models and measuring the damage, so red-teamers and guardrail developers have something to test against.Read the thread Shipped · research use
- 04 · QuantizationQuantization & efficient servingBit-width work that turns large models into something a laptop or a single consumer GPU can hold.Read the thread Shipped · open weights
- 05 · In progressDetection models & reconfigurable siliconCompact models that inspect every request, and the longer job of compiling them onto reconfigurable chips so inspection stops costing what it costs today.Read the thread Actively researching
- 06 · GenomicsPopulation genomics for the GulfNo genomic foundation model has been built for Arab or Emirati populations. That absence is the gap, and it is the one we are working on.Read the thread Actively researching
How we work
What we hold ourselves to
Publish the weights, then the method.Open weights on the Hub from day one, with the write-up following once the evaluation is broad enough to support the claim. Shipping first is not a substitute for measuring — it is an admission of what we have measured so far.
Report the number that could embarrass us.Refusal rate without KL divergence, or WER without parameter count, tells you the half of the story that flatters the model. Both halves or neither.
Efficiency is a research direction.A model that cannot be served on the hardware someone owns is a result, not a tool. Compression and quantization are the thread every other thread depends on.
Document what failed.Negative results and failure modes go in the write-up. A defence you cannot verify is a defence you cannot trust.