A 27B-class model now runs on a phone — Bonsai 27B clears the mobile memory ceiling
PrismML (founded by Caltech academics) released Bonsai 27B on 2026-07-14 — 1-bit and ternary (1.58-bit, {-1,0,+1}) builds of Qwen3.6-27B, Apache-2.0. The ~3.9 GB 1-bit build (~14× smaller, ~90% of full quality) is the one that clears the iPhone 17 Pro memory ceiling; the ~5.9 GB ternary build (~9× smaller, ~95%, ~58 tok/s on an M5 Max) targets laptops. Capable AI can now live fully on-device, offline, with no data leaving the phone.
Why it matters (spine — atoms). Doing more with fewer bits, out to the edge. Spec an offline, private on-device assistant for the safety-sensitive gardens (a child-safe tutor, a rural agronomy copilot, a clinic tool) where 'nothing leaves the device' is the feature.
What would prove it wrong (falsifier). Independent testing shows the on-phone build is materially degraded (well below ~94%) or unusably slow for the target task.
Live-status check + triangulation. PrismML (founded by Caltech academics) released Bonsai 27B on 2026-07-14 — 1-bit and ternary (1.58-bit, {-1,0,+1}) builds of Qwen3.6-27B, Apache-2.0. The 3.9 GB 1-bit build (14× smaller, 90% of full quality) is the one that clears the iPhone 17 Pro memory ceiling; the ~5.9 GB ternary build (9× smaller, ~95%, ~58 tok/s on an M5 Max) targets laptops. Capable AI can now live fully on-device, offline, with no data leaving the phone. Source: Company · PrismML (Caltech) (2026-07-14); see source link. Every post-cutoff figure was tool-fetched against a live source during the Kimi K3 sweep (2026-07-20); see sweeps/2026-07-20-VERIFIED_DIGEST.md for the full provenance and any corrections of record.
Spec an offline, private on-device assistant for the safety-sensitive gardens (a child-safe tutor, a rural agronomy copilot, a clinic tool) where 'nothing leaves the device' is the feature.
domains: edge-compute · quantization · atoms