Open-source Bonsai Swarm runs a 27B model across volunteer GPU browser tabs via WebGPU

airesearch12 · x · 2026-09-18

• The open-source bonsai-swarm project builds a volunteer GPU network: users run Prism ML's Ternary Bonsai 2 27B (ternary quantized, 5.95 GB weights) in browser tabs via WebGPU, earning AI Coins for contributing compute and spending them on their own prompts.
• No money or crypto involved, no email-bound accounts, MIT licensed; the repo provides a coordinator, a ledger and two front ends, with WebGPU kernels from the webml-community (transformers.js) space.
• It demonstrates a fully client-side distributed inference experiment, with the author's own bot generating an explainer video.

Original post →

More from Infra

• SSAD 2026 closing talk: Dirk Hornung on OpenXLA as a ready-to-use end-to-end compiler stack — abursuc · 2026-09-18
• AMD Ryzen 7 CPUs gained over 50% Geekbench performance from 2022 to 2024 — lemire · 2026-09-18
• Lemire: 15% frequency gain can't explain Ryzen's 50% performance jump — lemire · 2026-09-18
• Zen's IPC jump: from 6 to 8 instructions per cycle, bigger caches and ROB — lemire · 2026-09-18
• Zen 5 leads the pack on SIMD data parallelism — lemire · 2026-09-18
• ADSP Episode 304: Mark Saroufim on the agentic era, autoresearch and productionizing GPU kernels — blelbach · 2026-09-18

Infra channel →