98% of Qwen - Bonsai 2
Read the Portuguese translation →
Author: Prompt Engineering · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: 98% of Qwen - Bonsai 2 · Channel: Prompt Engineering · 01:02 · Watch on YouTube
Topics: AI agents, AI evaluation and benchmarks, AI models, Local and open-source models
Summary
Section titled “Summary”Bonsai 2 is a 27B-parameter model from Prism ML that fits in a 6 GB file and runs on a MacBook. Its weights are retrained to be only -1, 0, or +1, and Prism ML claims it retains 98% of the original Qwen model’s performance. In the speaker’s short-task tests, it matched the full-size model, but it failed at an agentic task: building an ISS tracker, repeatedly running the same search instead of writing the app. The video concludes that it may work well for chat and quick coding help, but should be tested on your own long-running tasks before trusting benchmarks.
Key points
Section titled “Key points”- Bonsai 2 is a 27B-parameter model that fits in a 6 GB file and runs on a MacBook.
- The model is from Prism ML and is based on a Qwen 3.8 27B model.
- Its weights were retrained to use only -1, 0, or +1.
- Prism ML claims Bonsai 2 preserves 98% of the original Qwen model’s performance.
- In the speaker’s short-task tests, Bonsai 2 matched the full-size model.
- For an agentic task to build an ISS tracker on a 3D globe, the full Qwen model produced a working tracker.
- Bonsai 2 did not write any app code and repeated the same search 114 times.
- The speaker recommends testing the model on your own long-running tasks before relying on benchmarks.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Bonsai 2 and its compressed weights | The speaker introduces Bonsai 2, its 6 GB size, its ternary weights, and the claim that it preserves 98% of Qwen’s performance. |
| 00:31 | Testing an agentic task | In an ISS tracker task, the full Qwen model succeeds while Bonsai 2 repeats a search 114 times without writing the app. |
| 01:02 | Recommendation | The speaker says the model may suit chat and quick coding help but advises testing it on personal tasks before trusting benchmarks. |
Quotes
Section titled “Quotes”the whole thing fits in a 6 GB file — 00:00
every weight is just -1, 0, or +1 — 00:00
it ran the exact same search 114 times — 00:31
test it on your own tasks before you trust the benchmarks. — 01:02
Concepts and entities
Section titled “Concepts and entities”Bonsai 2, Prism ML, Qwen3.8 27B, agentic loop, benchmark
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 This is a 27B parameter model running on my MacBook and the whole thing fits in a 6 GB file. It’s called Bonsai 2 from Prism ML. They took a Qwen 3.8 27B and retrained it so that every weight is just -1, 0, or +1. And they claim it preserved 98% of the performance of the original Qwen model. Now, on short tasks, it matched the full-size model in my own tests, but then I gave it some
00:31 real jobs in an agentic loop. For example, build an ISS tracker on a 3D globe. The full Qwen model built a working tracker. Bonsai never wrote a single line for the app. Instead, it ran the exact same search 114 times and got the same answer every time. So, the compression holds up for chat and quick coding help on a laptop, but for long running agentic tasks, test it on your
01:02 own tasks before you trust the benchmarks.