Qwen3.6 27B versus 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.

Sources

• Qwen3.6 27B base model card
• PrismML Bonsai 27B documentation
• PrismML Bonsai 27B whitepaper
• PrismML formats and runtime support

Related Bonsai 27B lessons

• What is Ternary Bonsai 27B?
• How good is Ternary Bonsai 27B?
• Why Ternary Bonsai 27B is called a 1.58 bit model

The benchmark numbers on this page describe one checkpoint, runtime, machine, and test shape. Reproduce the test on the hardware and workload you plan to use before making a product decision.