Toronto AI-compression company whose ShapeLearn technology learns optimal per-parameter numeric precision to shrink and accelerate large AI models, built on a University of Toronto research group's decade of quantization research.
Canadian innovation builds a brighter tomorrow.
Toronto AI-compression company whose ShapeLearn technology learns optimal per-parameter numeric precision to shrink and accelerate large AI models, built on a University of Toronto research group's decade of quantization research.
Federally incorporated as ByteShape Inc. on 2024-11-28, status Active, with a matching Ontario registration; the registry lists the registered office in Ottawa. ised-isde.canada.ca
Gives its own location as Toronto and describes itself as a University of Toronto spinout — the registered office in the federal registry is an Ottawa address, not the operating site. byteshape.com
A University of Toronto spinout: ShapeLearn is built on Prof. Andreas Moshovos's ECE research group and co-founder Milos Nikolic's PhD research on model quantization. twosmallfish.vc
Backed by Two Small Fish Ventures, an early-stage deep-tech VC fund, and incubated by University of Toronto Early Stage Technology (UTEST). utest.to
ShapeLearn compresses a 35B model to 3.8x smaller weights (4.19 vs 16 bits/weight) and runs it 2.6x faster on an RTX Pro 6000 versus the original BF16 version. byteshape.com
Compression retains 99.27% of original model benchmark performance while enabling a 35B model to run on a 24GB consumer GPU (RTX 4090) at up to 285.53 tokens/second. byteshape.com
ShapeLearn's technique is grounded in the group's own peer-reviewed papers, including GOBO (MICRO 2020), Mokey (ISCA 2022) and Schrodinger's FP (MLSys 2024). byteshape.com
Founder Andreas Moshovos is a University of Toronto Electrical and Computer Engineering professor and Vector Institute faculty affiliate. ece.utoronto.ca
Products ByteShape publishes on its own site. Each card links to the page it was taken from.
Model compression technology
Proprietary optimization technique that learns bit-length and numeric-format allocation across a model's components to hit a target hardware, memory and quality budget, validated on-hardware before deployment.
Learn morePre-compressed open models
Compressed versions of open models (Qwen 3.6-35B, Qwen3 Coder 30B, Qwen 3.8 27B vision-language, Qwen-Image-2512, Krea-2-Turbo) distributed in GGUF and other formats for llama.cpp, LM Studio, Ollama, ComfyUI and vLLM-Omni.
Learn moreHow ByteShape says its technology can be deployed and delivered.
Sourced from the company's own pages or announcements.
Andreas Moshovos
Founder & CEO
Enrique Torres Sanchez
Founder
Milos Nikolic
Co-Founder
Ali Hadi Zadeh
Co-Founder
Terms the company uses about its own work.
Other companies in AI.
Discover companies, track funding, explore career opportunities and stay informed about Canada’s growing tech ecosystem.
