Caltech spin-out PrismML raises $16M and open-sources 1-bit AI model claiming 8x speed gains

What's the deal? PrismML, a startup founded by Caltech researchers, has emerged from stealth with a 1-bit large language model it claims achieves radical compression without sacrificing reasoning performance. The company raised $16.25 million in a SAFE and seed round fromKhosla Ventures, Cerberus CapitalDealroom has a profile for this one. Try Dealroom →, and CaltechDealroom has a profile for this one. Try Dealroom →. It has simultaneously open-sourced its flagship model, Bonsai 8B, making it freely available to developers and researchers.

The core claim: a standard 16-bit model with a 16 gigabyte memory footprint can be compressed to 1 gigabyte at 1-bit precision, with processing speeds up to eight times faster and energy consumption reduced by 75–80% on current hardware.

Why now? AI's energy and compute costs have become a serious constraint. Data centre build-outs are straining power grids, and the economics of running large models on consumer devices — phones, laptops, wearables — have remained impractical. PrismML argues its approach changes that calculus, enabling on-device AI that was previously impossible due to memory and energy limitations.

Investor Vinod Khosla called it "a mathematical breakthrough, not just another tiny model" — a distinction that matters because PrismML's framework can theoretically be applied to any model architecture, including transformers and diffusion models.

What could go wrong? The claims are extraordinary and come entirely from the company itself. Independent benchmarking of the Bonsai 8B model — now that it is open-sourced — will quickly test whether the performance holds up across diverse tasks. Model compression often involves trade-offs that show up in specific domains, and "comparable to 16-bit models" is a claim that needs rigorous external validation.

The intellectual property is owned by Caltech, with PrismML as the sole exclusive licensee — a structure that creates dependency on the university relationship and could complicate future fundraising or commercial agreements.

The signal: PrismML joins a growing wave of research-to-startup ventures betting that the next phase of AI will be won not by building bigger models but by making existing ones far more efficient. If 1-bit compression delivers on its promise, the implications are significant: cheaper inference, longer device battery life, private on-device processing without cloud dependency, and more accessible AI in low-bandwidth or low-power environments. The open-source release is a deliberate move to accelerate adoption and stress-test the technology in public — the results will quickly tell the market whether this is a genuine breakthrough or an overstated claim.

Source:
WSJ

Image Credit:
PrismML

A.M.