---
title: "Bonsai 27B Runs Qwen 3.6 27B at 10x less memory."
source_url: https://youtu.be/V6LmF7TuBmY
captured_at: 2026-07-15
author: "Tim Carambat"
contributor: "ชิดชนก"
platform: youtube
---
# Bonsai 27B Runs Qwen 3.6 27B at 10x less memory.

youtube video by **Tim Carambat**, 17m39s, uploaded 2026-07-14. URL: https://youtu.be/V6LmF7TuBmY

## Shared in LINE Note+
- 2026-07-15 11:38 by ชิดชนก

## Description
The new Bonsai 27B Model from PrismML is Qwen3.6 27B - a beloved workhorse for many - into something you can run on local computer.

10x less memory and a much more modest file size while still benchmarking 95% of the original FP16 model. Of course, no model is perfect, but if you have been wanting to run Qwen3.6 27B and dont have the computer or headroom - you finally can.

The Ternary format is much smarter, but takes a bit more compute. The Binary model is enough to fit into a phone form factor. Imagine 27B, even remotely, in your pocket. That is now a reality.

**Links** :
Blog: 
Whitepaper: 
HuggingFace Collection: 
WebGPU Demo: 

**As of now you MUST use these forks to run these models**

PrismML LLama.cpp repo: 
MLX Repo: 

AnythingLLM: 
OpenComputer: 

**Chapters** :
0:00 The promise of Bonsai 27B
0:57 Lets look at the benchmarks and models
2:52 Why would you choose Bonsai 27B?
5:22 How to run it locally...
6:00 Download the right binary
7:27 Download the model from HuggingFace
7:50 (Shortcut) - Use webGPU!
8:17 Picking the right files from HuggingFace
9:00 Start the llama-server and use the UI
10:52 How does it work with simple agent tasks?
12:12 What about using a whole c