Prism ML rolled out Bonsai 27B, a version of the Qwen 3.6 large-language model that fits on a mobile phone. By squeezing the 54 GB original down to under 4 GB with 1-bit quantization, the company delivers a 14× size reduction and lets users run the model locally, without cloud API calls.

Why the compression matters

LLMs usually need a desktop-class GPU or a paid API because their weights take up dozens of gigabytes. Quantization—using fewer bits per weight—shrinks the model but can also strip away knowledge. Bonsai offers two extremes: a ternary variant (three possible weight values, 7.2 GB) and an aggressive 1-bit variant (3.9 GB). The trade-off between size and capability drives the test.

How the models fare on factual recall

The original Qwen 3.6 answered every historical-date question correctly in the tester’s suite. The ternary Bonsai missed five of six, and the 1-bit version failed every single date query. As expected, aggressive compression discards rare, specific facts first, preserving only the broad language patterns that drive fluency.

Coding performance tells a different story

When the prompts switched to coding tasks, the results flipped. All three models solved classic concurrency bugs without error. On a demanding React animation challenge, the 1-bit Bonsai finished in two attempts, while the original needed four because it spent extra time parsing the code. The ternary version required ten attempts and eventually crashed the test server.

Practical hurdles

Bonsai does not run on the popular Ollama runtime. The 1-bit model needs a custom execution layer supplied by Prism ML, so users must install non-standard software to get it working. This dependency limits the model’s appeal to those comfortable tweaking their environment.

Who should consider Bonsai, and who should stay away

  • General-purpose chat – The drastic loss of factual detail makes Bonsai unsuitable as a knowledge-base assistant. For accurate answers to history, science, or current events, stick with the original model or a cloud API.
  • Local coding aide – Developers who want an offline tool that can suggest code, catch bugs, and run on a phone or low-end laptop will find the 1-bit version delivers speed and low storage cost. It is not an autonomous agent, but it handles logic and syntax efficiently.

Bottom line

Bonsai 27B shows you can compress a 54 GB LLM to a phone-friendly 3.9 GB and still get surprisingly fast, competent code suggestions. The price is a near-total erosion of specific factual recall, so the model belongs in the toolbox of developers who value offline speed over encyclopedic knowledge.