LabAn AI research lab · six threads
We build our own models. And the silicon to run them.
Open-weight models compressed for real deployments, and a roadmap that takes them from software into reconfigurable chips.
46open-weight models
506kall-time downloads
40kdownloads this month
What we research
Six threads, from expert pruning through to population genomics. Each has its own page with the detail and the numbers behind it.
- CompressionExpert pruningCutting experts out of large mixture-of-experts models until they fit the hardware people actually own.Shipped · open weightsRead the thread
- SpeechMultilingual speech recognitionDialect-aware recognition for languages where labelled audio is scarce and the incumbent systems are far larger than they need to be.Shipped · leaderboard-tierRead the thread
- EvaluationRefusal ablationRemoving refusal behaviour from open models and measuring the damage, so red-teamers and guardrail developers have something to test against.Shipped · research useRead the thread
- QuantizationQuantization & efficient servingBit-width work that turns large models into something a laptop or a single consumer GPU can hold.Shipped · open weightsRead the thread
- 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.Actively researchingRead the thread
- 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.Actively researchingRead the thread
What we have shipped
Open weights on Hugging Face from day one. Model cards are readable on this site — you only leave to download.
Refusal ablation
Refusal-ablation research, measured
- Gemma-4-12B-uncensored-bf161.3k
- Qwen3.6-27B-V2-abliterated-uncensored-Q4_K_M-GGUF28k
- Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-OptiQ-3.7bpw-mlx32k
Quantization
Builds that fit the hardware you have
- DeepSeek-V4-Flash-TQ-Q5.6-MLX4.5k
- Step-3.7-Flash-OptiQ-3.7bpw-mlx2.5k
- Keye-VL-2.0-30B-A3B-uncensored-mlx-optiq-3.7bpw731
Where it is going
One timeline, starting with what already exists. Compact models today, inline inspection next, reconfigurable silicon after that.
- ShippedNow
- Open weights on the Hub
- Expert pruning
- Speech for underserved languages
- Actively researchingNext
- Detection models that run inline
- More languages
- A genomics benchmark for the Gulf
- DirectionLater
- Onto reconfigurable silicon
- A deployable appliance
- A population model for the Gulf
Work with us
Design partners, research collaboration, and national genomics programmes that need the model to come to the data.