# Bonsai 27B: A 27B-Class Model That Runs on a Phone

AI in your pocket: 27B-parameter model comes to mobile devices.

- **Source:** [Hacker News](https://news.ycombinator.com)
- **Published:** 2026-07-14
- **Topic:** [Data & AI](https://brifing.si/c/data-ai)
- **Original article:** <https://prismml.com/news/bonsai-27b>
- **On Brifing:** https://brifing.si/s/592bd6c3-736c-4c91-9466-0e926dea1352

## A Revolutionary Model

Bonsai 27B, based on Qwen3.6 27B, is introduced as the first 27B-class multimodal model capable of running on phones. Previous versions proved that commercially useful models could be created with 1-bit and ternary weights.

## New Capability Frontiers

This model offers new capability tiers including multi-step reasoning, structured tool calls, vision tasks, and computer-use agentic loops that remain coherent across many steps. Previously, running such a model locally was impractical.

## Two Distinct Variants

Bonsai 27B comes in two variants: Ternary Bonsai 27B is the quality-oriented version at 5.9 GB, running on everyday laptops. 1-bit Bonsai 27B is the footprint-oriented version at 3.9 GB, fitting within an iPhone 17 Pro's memory.

## Multimodal Support

Both variants support multimodal inputs. The vision tower ships in compact 4-bit form, enabling on-device processing of screenshots, documents, and camera input. The model also carries a full 262K-token context.

## Performance Comparison

Across 15 benchmarks, Ternary Bonsai 27B retains 95% of full-precision performance, while 1-bit Bonsai 27B retains 90%. Critical areas like math and coding show nearly untouched performance.

## Intelligence Density Breakthrough

1-bit Bonsai 27B delivers 0.53 intelligence per GB - 10x the full-precision baseline and 2.7x better than the best low-bit alternative. This packs 27B-class capability into less space than a full-precision 2B model.

## The Importance of Local Execution

On-device operation keeps user data local and enables zero marginal cost for hundred-step loops. This unlocks persistent on-device agents, offline assistants, and systems that reason over private local data by design.

## Sources covering this story

- https://prismml.com/news/bonsai-27b — 

---

_Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._
