Ternary Bonsai 2 27B
Ternary Bonsai 2 27B is a compressed build of Qwen3.8-27B, released in September 2026 by PrismML under the Apache 2.0 license. Every weight is squeezed down to almost nothing (so-called ternary compression), so the 27-billion-parameter model fits in 5.95 GB (PTQ1_0 format) or 7.21 GB (PQ2_0), instead of 53.8 GB at full precision. It reads 262,000 tokens of context. Across 14 thinking-mode benchmarks, PrismML reports an average of 84.78, against 86.32 for the uncompressed original, 72.59 for the classic IQ2_XXS quantization (9.4 GB) and 85.18 for UD-Q4_K_XL (17.6 GB).
Strengths
- A 27-billion-parameter model in 5.95 to 7.21 GB, against 53.8 GB for the original version
- 84.78 average across 14 thinking-mode benchmarks, against 72.59 for the classic IQ2_XXS quantization, which weighs more (9.4 GB)
- 262,000 tokens of context and an Apache 2.0 license that allows commercial use
Limitations
- Stock llama.cpp cannot read these files: you have to build the PrismML fork, and Ollama and LM Studio stay out of the game meanwhile
- All scores come from the publisher, no independent measurement published so far
- The Ollama command shown on Hugging Face downloads the 53.8 GB full-precision version, not the compressed one
Best for
- Developers with an 8 GB GPU who want a local model sturdier than a 7B
- Teams that have to keep their data on their own machines and are fine building a tool from source