Bonsai 27B field guide
Qwen3.6 27B versus Ternary Bonsai 27B
Compare the base Qwen3.6 27B model with PrismML's ternary representation across architecture, size, quality, runtime, and use cases.
What is the difference between Qwen3.6 27B and Ternary Bonsai 27B?
Qwen3.6 27B is the base multimodal model. Ternary Bonsai 27B keeps that architecture but changes the language weight representation and runtime path. Qwen's full precision language weights need about 54 GB, while PrismML reports a 5.9 GB ideal ternary representation and about 7.2 GB for its current deployed GGUF language model.
What stays the same
- The language model has 64 blocks and a 262,144 token context.
- The architecture mixes linear attention and full attention.
- The model accepts text and images and returns text.
- The Qwen3.6 training and post training behavior is the starting point.
What changes
PrismML maps the language weights into ternary values and applies group scales inside low bit kernels. The vision tower remains a separate 4 bit component. The deployment uses PrismML compatible MLX, llama.cpp, or CUDA paths instead of a normal FP16 runtime.
The smaller representation reduces memory traffic but also changes benchmark results. PrismML reports 80.49 for ternary against 85.07 for its FP16 Qwen reference.
Questions people ask
Can I use the normal Qwen Transformers command with the ternary model?
Use the runtime and format described by PrismML. The low bit files need kernels that understand their packing.
Does Ternary Bonsai change the context limit?
PrismML documents the same 262,144 token maximum, though available memory can set a lower practical limit on a local machine.
Primary sources
Vendor claims on this page are labeled as PrismML claims. Local results are added only after a saved trace completes.