PrismML says Ternary Bonsai 2 27B is based on Qwen3.8 27B, has a stated 5.9 GB footprint, is nine times smaller than its full-precision counterpart and is offered under Apache 2.0. In a post on X, @PrismML says the model retains 98.2% of the reference model’s aggregate benchmark performance.

The announcement presents the release as a quality improvement over the first Bonsai 27B model. PrismML says the footprint remains 5.9 GB while the gap to full precision has narrowed, highlighting gains in agentic coding, multimodal reasoning and long-horizon tool use. It does not explain the underlying ternary technique or what the stated footprint includes.

What the reported comparison shows

The comparison published with the announcement gives Qwen3.6 27B, Qwen3.8 27B and Ternary Bonsai 2 27B scores across six capability categories. It reports an overall score of 83.9 for Ternary Bonsai 2 27B, compared with 85.4 for Qwen3.8 27B. That produces the stated 98.2% retention figure.

Capability

Qwen 3.6 27B

Qwen 3.8 27B

Ternary Bonsai 2 27B

Retention to Qwen 3.8

Knowledge and reasoning

84.71

86.66

83.95

96.9%

Math

94.64

97.06

96.57

99.5%

Coding

82.57

82.17

81.58

99.3%

Agentic and tool calling

80.05

79.74

77.57

97.3%

Instruction following

74.53

81.25

82.66

102%

Vision

79.82

81.64

78.59

96.3%

Overall

83.6

85.4

83.9

98.2%

The announcement does not identify the benchmarks, test prompts, sample sizes or evaluation hardware, and it does not explain how the aggregate score and retention percentages were calculated. The figures should therefore be read as PrismML’s reported comparison, not as an independently verified evaluation.

Where the reported results are closest

On the published figures, math is the closest category to Qwen3.8 27B: Ternary Bonsai 2 27B scores 96.57 versus 97.06, or 99.5% retention. Coding is also close, with a reported 81.58 compared with 82.17 and 99.3% retention.

Instruction following is the only listed category where Ternary Bonsai 2 27B scores above Qwen3.8 27B. Its reported score is 82.66 versus 81.25, represented in the comparison as 102% retention. That result is specific to this category and evaluation; it does not establish that the smaller model is generally better.

The widest reported gaps are in vision and knowledge and reasoning. Ternary Bonsai 2 27B scores 78.59 versus 81.64 in vision, or 96.3% retention, and 83.95 versus 86.66 in knowledge and reasoning, or 96.9% retention. Agentic and tool calling is reported at 77.57 versus 79.74, corresponding to 97.3% retention.

What the announcement says about size and licensing

PrismML describes Ternary Bonsai 2 27B as nine times smaller than its full-precision counterpart and says its footprint is 5.9 GB. The announcement does not define what that footprint includes, so the figure should not be treated as a complete hardware or deployment requirement.

PrismML says the model is available under Apache 2.0. The announcement does not provide a download link, hardware requirements, inference-speed measurements, runtime support or deployment instructions, so the license and reported size alone do not establish how the model can be run in a particular environment.

What the figures do and do not show

The reported numbers support a narrower conclusion: PrismML is presenting Ternary Bonsai 2 27B as a substantially smaller model based on Qwen3.8 27B, with an overall score close to the reference model in the supplied comparison. The category results vary, with the smallest reported retention in vision and the largest relative result in instruction following.

They do not show how the model performs in a particular application or how it will compare under different evaluation conditions. A model card, benchmark methodology and independent testing would be needed to assess those questions beyond the announcement’s reported figures.

Sources