PrismML Bonsai 27B

A dated reading of what is claimed, reported, and independently verified in the current evidence.

Current reading

Readiness band and full integer triple

AI-assisted assembly · derived results

Readiness band

Watch

Primary summary from verified readiness

Evidence strength Critical

Decision

Human editorial judgment · 2026-07-21

Proceed with caution

Why: The core compression result (1.125 bits/weight, ~89.5% FP16 retention on a 27B model, independently written up by MarkTechPost) is real and notable, but upstream kernel support is still unmerged and the vendor itself discloses agentic-coding and vision-quality gaps larger than the headline retention number suggests. The candidate's original framing also slightly overstated the WebGPU/iPhone-15-16 deployment story relative to PrismML's own primary materials, though a linked (non-native, community-built) WebGPU demo does exist.
Next: Reproduce PrismML's own MLX benchmark on real hardware (M-series Mac or iPhone) before citing throughput numbers — verified generation speed is only ~11 tok/s on an iPhone 17 Pro Max for the 1-bit build. When referencing this candidate, state the native runtime correctly as MLX/CUDA (not WebGPU) and the demoed hardware as iPhone 17 Pro Max (not iPhone 15/16 Pro); a community-built WebGPU browser demo is linked from PrismML's own page but is not the native path.

Constraints

No named blocker is present in the current public projection.

Evidence summary

AI-assisted assembly

Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.

Publication record

Initial public reading