Ternary Bonsai 2: 5.9GB Model Matches Qwen Performance

PrismML has shipped Ternary Bonsai 2 27B, a compressed rework of Qwen3.8 27B that fits in 5.93 GB instead of the original's 53.80 GB in FP16. The company says the model holds onto 98.2% of the parent's average score across 20 benchmarks, a jump from the roughly 95% retained by the first Bonsai 27B released two months ago. The model handles both text and images, supports a 262K-token context window, and PrismML has shown it running Cline coding agents and computer-use tasks on a single RTX 5090.

The weights are Apache 2.0 licensed and already runnable, PrismML says, on a 16 GB laptop or a 24 GB GPU, provided you're using its own llama.cpp fork or MLX runtime. The architecture itself is untouched from Qwen3.8 27B: 27.36B parameters split across a 24.35B backbone, 2.54B in embeddings and the LM head, and a 0.47B vision tower that loads separately as a 0.63 GB GGUF file only when an image shows up. The backbone runs hybrid attention, mixing linear and full-attention layers. The real story is how PrismML squeezed that footprint down in the first place, and it starts with how the ternary format itself represents each weight.