ByteShape

Company / Toronto, Canada

Make capable AI practical to deploy.

We are a University of Toronto spinout bringing machine learning and computer architecture together.

Invest or partner with us Company overview

Published engineering

A 35B model.
One 24 GB GPU.

Our Qwen3.6-35B-A3B build brings a large language and vision model to a desktop graphics card.

Explore the deployment

285.53

tokens per second

99.27%

of its full-precision benchmark score

RTX 4090 (24 GB), with multi-token prediction.

Read the published result (opens in a new tab)
Measurement details

Qwen3.6-35B-A3B · MTP-GPU-5 · RTX 4090 (24 GB) · llama.cpp with multi-token prediction. Mixture of experts: 35B total parameters, 3B active. Model file: 18.61 GB, separate from total runtime memory.

Benchmark score relative to this model's original BF16 baseline, using the release's evaluation suite.

Our direction

Connect AI software
and hardware.

Our long-term goal is a common layer for how model data is stored, moved, and used in computation.

Public model releases

Language, vision-language, and image-generation models to run today.

Browse the catalog

Team

Built at the intersection of AI and computer architecture.

Researchers and engineers in learned quantization, software, and the hardware that runs AI.

  • Co-Founder Andreas Moshovos

    Andreas Moshovos

    Co-Founder & CEO

    Professor of Electrical and Computer Engineering at the University of Toronto. Designs high-performance processors and memory systems, with work that has influenced commercial designs.

    Faculty page
  • Co-Founder Milos Nikolic

    Miloš Nikolić

    Co-Founder & Lead ML Acceleration

    Develops learned data representations for efficient machine learning. PhD in Computer Engineering, University of Toronto.

    LinkedIn
  • Co-Founder Enrique Torres Sanchez

    Enrique Torres Sanchez

    Co-Founder & AI Systems Lead

    Leads AI systems development. Brought ShapeLearn from research into a product. PhD candidate in Computer Engineering at the University of Toronto.

    LinkedIn
  • Co-Founder Ali Hadi Zadeh

    Ali Hadi Zadeh

    Co-Founder & Lead Scientist

    Researches quantization and AI efficiency, and leads the evaluation and benchmarking of optimized models. PhD in Deep Learning Acceleration, University of Toronto.

    LinkedIn
  • Kareem Ibrahim

    Kareem Ibrahim

    Machine Learning Engineer

    Works on vLLM, quantization methods, and model analysis. PhD candidate in Computer Engineering at the University of Toronto.

    LinkedIn

Advisors

Advised by founders, investors and researchers.

Operators who built and scaled technology companies, and professors in computing and machine learning.

  • Advisor Eva Lau

    Eva Lau

    General Partner & Co-Founder, Two Small Fish Ventures

    Engineer and deep-tech investor. Founding team member at Wattpad, where she helped scale the platform to tens of millions of users.

    Two Small Fish profile (opens in a new tab)
  • Advisor Allen Lau

    Allen Lau

    Operating Partner & Co-Founder, Two Small Fish Ventures

    Co-founder and former CEO of Wattpad, which grew to 100 million users and was acquired by Naver. Board member at MaRS Discovery District.

    Two Small Fish profile (opens in a new tab)
  • Advisor Dan Trepanier

    Dan Trepanier

    MSFA Director & Lecturer, Santa Clara University

    Founder and former CEO of Quake Technologies, acquired by Applied Micro Circuits. Former CEO of Xambala. Teaches data science and finance at the Leavey School of Business.

    SCU faculty page (opens in a new tab)
  • Advisor Nick Koudas

    Nick Koudas

    Professor of Computer Science, University of Toronto

    Researches large-scale data management and applied machine learning. Co-founded Sysomos (now part of Meltwater) and Aislelabs (acquired by Constellation Software), and is a co-founder at Workorb. University of Toronto Inventor of the Year 2011.

    Faculty page (opens in a new tab)
  • Advisor Kunle Olukotun

    Kunle Olukotun

    Professor of Electrical Engineering and Computer Science, Stanford University

    Pioneer of multicore processor design and leader of the Stanford Hydra project. Founded Afara Websystems, acquired by Sun Microsystems, and co-founded SambaNova Systems. Member of the National Academy of Engineering.

    Stanford profile (opens in a new tab)

Research origin

From how bits are stored to what models can do.

Our team's published work connects learned numerical representations, efficient AI hardware, and model evaluation.

Our evaluation methodology (opens in a new tab)

Our investors

Open the one-pager (PDF) (opens in a new tab)

Start a conversation.

Contact our team about partnerships, investment, or your next deployment.