35B on a 24 GB GPU
285.53 tok/s
Qwen3.6 on RTX 4090 with multi-token prediction.
Inspect the results (opens in a new tab)Custom optimization
Model optimization.
Built into your workflow.
ShapeLearn is ByteShape's automated optimization library and SDK. It learns the precision your model needs to meet your hardware and quality targets.
Choose how to
use ShapeLearn.
How we evaluate models (opens in a new tab)
- 01
Plan SDK integration
Independent SDK use is our planned product path. Talk to us about bringing ShapeLearn into your own model pipeline.
- 02
Build alongside us
Apply ShapeLearn with our team, from integration to optimized model files, benchmark results, and runtime settings.
See it in our
published models.
30B on a Raspberry Pi
8.03 tok/s
Qwen3 on Raspberry Pi 5 (16 GB), running on the CPU with llama.cpp.
Inspect the results (opens in a new tab)Smaller image-model weights
5.26× smaller
Qwen-Image-2512: 40.86 GB to 7.77 GB of diffusion weights.
Compare the images (opens in a new tab)Bring us your next model.
Tell us what you want to optimize and how you'd like to work together.