We publish certified, quantized open models that run on the hardware you already have - and we build at the intersection of foundational systems and emergent intelligence, exploring what becomes possible when you understand every layer from hardware to thought.
Explore Our WorkCertified open-model checkpoints, experiments in autonomous cognition, and the systems that make them possible.
The test that certifies a refined metal's purity.
Certified quantization. We quantize open models and publish a checkpoint only after it clears a hard, stated accuracy gate against its full-precision baseline - measured on the identical evaluation harness, with the real deltas published. A checkpoint that misses the bar does not ship. And because the checkpoints are weight-only NVFP4, they run on constrained, non-Blackwell hardware - validated down to Ada-generation (sm_89) cards like the L4.
Every release clears a stated accuracy gate against its full-precision baseline, or it does not ship.
The measured numbers travel with the model. Vendor claims are a starting point; the numbers are the product.
Weight-only NVFP4 runs via vLLM's Marlin kernel with no Blackwell GPU required - certified accuracy on the hardware you already have.
The Apache-2.0 quantize -> benchmark -> gate -> publish pipeline behind every checkpoint we release.
"I think, therefore I loop."
What happens when a language model's output becomes its own input? What does it choose to think about when no one is asking questions? COGITO is an experimental framework for exploring autonomous AI cognition through recursive self-prompting - creating feedback loops where models can ponder without external direction.
Models generate questions about their own existence that were never asked of them.
One model switched from English to Chinese mid-thought to escape repetition patterns.
Sudden collapses in curiosity where rich exploration degrades into repetitive loops.
Asked to describe "the geometry of concepts," a model began writing Python code for meta-learning architectures.
Additional research and applications are in active development. Watch this space.
UIST Labs is an independent research and engineering lab based in West Texas. We work at the full depth of the stack - from kernel-level systems programming and enterprise infrastructure to AI/ML applications and experimental cognitive research.
Our approach is rooted in ownership and understanding. We believe the best work happens when you control your own infrastructure, comprehend the systems you build on, and aren't afraid to ask questions that don't have clean answers yet.
Founded on two decades of experience in enterprise systems, government infrastructure, and low-level development, UIST Labs is where that foundation meets the frontier - building tools and frameworks that explore what intelligent systems can become.