PrismML Releases Ternary Bonsai 2 27B: A 5.9 GB Apache 2.0 Model Retaining 98.2% of Qwen3.8 27B Performance
PrismML released Ternary Bonsai 2 27B, a 5.93 GB Apache 2.0 ternary-weight model retaining 98.2% of Qwen3.8 27B's benchmark performance.
PrismML released Ternary Bonsai 2 27B, a ternary-weight version of Qwen3.8 27B that shrinks the model from 53.80 GB FP16 to 5.93 GB under Apache 2.0 licensing. It scores 83.9 across 20 benchmarks versus 85.4 for the FP16 parent, and outperforms an IQ2_XXS quantization on AIME26 (95.83 vs 78.6) and LiveCodeBench v6 (90.07 vs 70.05). The 27.36B-parameter model supports text and images with a 262K-token context and runs on a 16 GB laptop or single 24 GB GPU via PrismML's llama.cpp fork or MLX runtime. Long-horizon agentic results retain only about 75% of full-precision scores, and all benchmarks are PrismML's own, not independently reproduced.
- 5.93 GB ternary model, roughly 9.1x smaller than the 53.80 GB FP16 baseline.
- Scores 83.9 on 20 benchmarks versus 85.4 for Qwen3.8 27B FP16.
- Beats IQ2_XXS quantization: 95.83 vs 78.6 on AIME26.
- Agent benchmarks retain only ~75% (SWE-bench Verified 60.8 vs 80.6).
- Requires PrismML's llama.cpp fork; stock llama.cpp cannot load the GGUF files.
- AI models
- Ternary Bonsai 2 27BQwen3.8 27B
From the original article
4 min read at marktechpost.com
PrismML has released Ternary Bonsai 2 27B, a ternary-weight version of Qwen3. 8 27B. The language model occupies 5.
By Asif Razzaq · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.