Breaking

PrismML Releases Bonsai 27B, Compressing a 27-Billion-Parameter Model Small Enough to Run on an iPhone

July 15, 2026 at 22:15 EDT

Foundation Models
Open Source
Infra & Chips

PrismML on July 14, 2026 released Bonsai 27B, a pair of extreme low-bit quantizations of Qwen3.6 27B, with a 3.9GB, 1-bit build that the company says is the first model of its size to run on a smartphone — specifically the iPhone 17 Pro at roughly 11 tokens per second.

July 14, 2026 · PrismML

A 27B Model That Fits in Your Pocket

Bonsai 27B compresses Qwen3.6 27B into a 3.9GB, 1-bit build that PrismML says is the first model of its size to run on a smartphone — roughly 11 tokens/sec on an iPhone 17 Pro.

3.9GB

1-bit build (1.125 bits/weight) — runs on a phone

5.9GB

Ternary build (1.71 bits/weight) — aimed at laptops

262K

Token context window, plus vision + tool calling retained

The memory math

Same 27B model, drastically smaller footprint. Column heights are proportional to memory required.

54GB

FP16

Full precision

~18GB

4-bit

Aggressive quant

5.9GB

Ternary

3.9GB

1-bit (phone)

Quality retained vs FP16

Across 15 self-reported thinking-mode benchmarks.

100%

FP16 baseline

~95%

Ternary

~90%

1-bit

Early praise

Clearly beats typical Q2, approaches Q4 quality at smaller size
1-bit build runs on modest PCs — 8GB RAM, no GPU
Coding and game-generation tasks described as "loop-proof"

Caution

Benchmarks are self-reported — need independent verification
Perplexity remains worse than the base model
Whether 90–95% claims hold under scrutiny will shape adoption

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.