PrismML Releases Ternary Quantized Model Bonsai 2 27B, Hits 98.2% of Base Model Performance in 5.9GB

AI lab PrismML has released a ternary quantized version of Bonsai 2 27B, upgrading its base model to Qwen3.8 27B with improvements in reasoning, coding, vision, and long-horizon agent capabilities.

The model scores 98.2% on comprehensive benchmarks while using less than one-ninth of the base model's memory, outperforming Qwen3.6 27B overall. It is only 5.9GB in total size, with 1.76 effective bits per weight, and supports a 262K context window.

On NVIDIA GeForce RTX 5090, it reaches 143TPS token throughput; on Apple Silicon M5 Max, it hits 46.8TPS. Its energy efficiency on RTX 4090 is 0.714mWh/Token, 40% lower than a full-precision 8B model.

The model runs on NVIDIA GPUs via CUDA and on Apple devices via MLX, with open weights under the Apache 2.0 license.