PrismML launches ternary bonsai 2 27b: a compact model with near-original performance

PrismML has unveiled Ternary Bonsai 2 27B, a ternary-weighted variant of Qwen3.8 27B, which retains 98.2% of its performance while occupying just 5.93 GB. The model supports text and image inputs, with a 262K-token context window, and runs on a 16 GB laptop or a single 24 GB GPU using PrismML’s llama.cpp fork or MLX runtime. It maintains the Qwen3.8 27B architecture, with 27.36B parameters split into language, embeddings, and vision components. The ternary weights cover most layers, with only 0.0976% in higher precision. PrismML reports that vision tasks retain 96.3% of performance, while knowledge and reasoning retain 96.9%. The model’s energy efficiency is 40% better than a full-precision 8B model. PrismML also highlights its performance on benchmarks like LiveCodeBench v6 and SWE-bench Verified, where it scores 90.07 and 60.8 respectively, compared to Qwen3.8 27B’s 70.05 and 80.6. The model is available in GGUF format, with separate vision tower files for image input.