PrismML released Ternary Bonsai 2 27B. It is based on Qwen3.8 27B. It retains 98.2% of the base model’s benchmark performance. It uses ternary weights with FP16 scaling for a 5.9GB footprint. It supports a 262K-token context window. It handles text and image input. It is licensed under Apache 2.0. It runs on NVIDIA GPUs and Apple devices. It reaches 143 tokens per second on an RTX 5090. It consumes 0.714 mWh per token on an RTX 4090. It improves reasoning, coding, vision and agentic performance over the first Bonsai 27B. It enables local coding agents, computer-use workflows and private document analysis.
What changed
Bonsai 27B retained 95% of full-precision performance in a 5.9GB footprint.
What it unlocks
Running a 27B-class model with near-full capability on local devices.
- 98.2% benchmark performance retained
- 5.9GB model footprint
- 262K-token context window
- 143 tokens/second on RTX 5090
What you need to act on it
- NVIDIA GPU via CUDA or Apple device via MLX
Sources