Your deployment sets the target.
The model, hardware, memory budget, and tasks you need it to perform.
Technology
Keep precision
where it matters.
AI models store billions of numbers. ShapeLearn learns which need more precision and which can use fewer bits, shaping the model around your hardware.
Different parts of the model. Different needs. Different precision.
Illustration of learned bit allocation.
Precision follows the model.
ShapeLearn learns bit lengths and numeric formats across the model.
The result runs on real hardware.
We measure task quality, speed, and memory, then deliver the selected build.
The allocation makes a difference
Better use
of every bit.
54.7% higher normalized benchmark score
Devstral Small 2 24B, at nearly identical generation speed on the same RTX 4080 (16 GB).
Optimize the execution, too.
Multi-token prediction takes Qwen3.6-35B-A3B from 214.54 to 285.53 tok/s on the RTX 4090 (24 GB), a 33.1% gain at the same benchmark score. GPU-5 and MTP-GPU-5, llama.cpp.
Explore MTP (opens in a new tab)From optimization to deployment
Built for the
tools you use.
Language and vision-language
GGUF builds for llama.cpp, LM Studio, and Ollama.
Find a build and run itImage generation
GGUF for ComfyUI and stable-diffusion.cpp. Humming for vLLM-Omni on Linux with supported NVIDIA GPUs.
Explore the runtimes (opens in a new tab)From model files to the hardware beneath them.
Our longer-term objective is a common layer between AI software and hardware: optimizing how tensors are stored and moved, independently of the formats used for computation.