Bonsai 2

Bonsai 2 is suitable for local chat assistants, knowledge Q&A and information retrieval, and lightweight coding assistance at the level of single code snippets on a laptop.

Bonsai 2 uses a full retraining starting from the Qwen3-27B architecture with ternary constraints as the objective, rather than applying post-training quantization to existing weights.

On simple tasks with thinking disabled and on long-context tests, Bonsai 2 performed nearly on par with full-precision Qwen3-27B, and the 98% claim largely held up.

Bonsai uses ternary quantization, compressing each weight to one of three values—-1, 0, or +1—with a shared scaling factor per 128 weights, requiring approximately 1.7 bits per weight, roughly 9 times smaller than 16-bit full precision.

The 98% score claimed by Bonsai 2 is the average measured across 20 benchmarks with the thinking feature set to maximum intensity.

Matching benchmark averages does not equal matching performance in real agentic work; the intelligence erased by compression is exposed in scenarios requiring long-range planning, self-checking, and sustained progress.

The Prism ML white paper reports that Bonsai 2 retains approximately 76% of the original model's score on TerminalBench 2.1 and approximately 75% on SWE-bench Verified, two long-chain benchmarks that were not included in the 98% average.

The Bonsai 2 model released by Prism ML claims to retain 98% of the performance of a 27-billion-parameter model within less than 8 GB of VRAM.

Bonsai 2 exhibited spinning-in-place behavior on long-chain tasks; for example, in the ISS tracker task it made 153 tool calls of which 150 were merely reading/listing/searching files, and it called the same tool 114 times.

Across six real-world long-chain agentic build tasks, Qwen3-27B scored 35/60 and delivered working applications, while Bonsai 2 collapsed across the board and produced no runnable pages whatsoever.

In an agentic task to build a 3D globe ISS tracker, the full Qwen model built a working tracker, while Bonsai 2 failed to produce a single line of code and repeated the exact same search 114 times.

Bonsai 2 is a ternary-quantized large model released by Prism ML, retrained from scratch on a Qwen3-27B base, with an overall file size of approximately 6 GB.

Bonsai 2 quantizes every weight in the model to one of three values—-1, 0, or +1—compressing each parameter to fewer than 2 bits.

Prism ML claims that Bonsai 2 retains 98% of the original Qwen model's performance after aggressive compression.

BitNet b1.58 restricts weights to the set {-1, 0, +1}, which is highly consistent with the approach taken by Bonsai 2.

In hands-on testing by a YouTube creator, Bonsai 2 performed nearly on par with the full-size Qwen model on short tasks (single-turn Q&A, conversation, one-shot questions).

Bonsai 2 is not suitable for long-running agentic tasks, complex workflows requiring multi-step planning, tool calls, and state retention, or automated processes with high demands on decision stability.

A single performance percentage is misleading; a model that claims to retain 98% of performance may happen to fall within the lost 2% for a specific task type, and real tasks should be tested before trusting benchmarks.

Bonsai 2 can run on consumer-grade laptops such as ordinary MacBooks, lowering the barrier to local inference for 27B-scale models.

PrismML's Ternary Bonsai 2 applies ternary quantization to the Qwen3.8 27B model, has received 3.87 million downloads, compresses the language model from 54GB to 5.9GB — a 9.3x reduction — and officially claims it retains 98.2% of the FP16 model's capability, with an average score of 84.78 on 14 reasoning benchmarks.