PrismML released Ternary Bonsai 2 27B on September 17, a compressed version of Qwen3.8 27B aimed at running reasoning and multimodal workloads locally. The company reports a 5.9GB model footprint, more than nine times smaller than the full-precision model, with weights released under Apache 2.0.

The model uses ternary weights—negative one, zero and positive one—with scaling factors. It supports text and image input, tool calling and a context window of up to 262,000 tokens. PrismML offers GGUF and MLX formats for local deployment.

Benchmark claims, not blanket parity

PrismML reports retaining 98.2% of aggregate benchmark performance against its full-precision baseline. Its published table gives Bonsai 2 an overall score of 83.9 versus 85.4 for Qwen3.8 27B, covering reasoning, coding, mathematics, instruction following, vision and tool use.

Those are company-reported evaluations, not an independent finding that every workload performs equivalently. Results vary by category: the compressed model trails the baseline on vision and tool-use scores while scoring higher on instruction following.

Deployment details matter

The release makes a larger model available within a smaller storage and memory envelope, but developers should distinguish the headline footprint from a complete installation. PrismML's demo downloads a larger packing by default, plus a vision projector and optional supporting software.

For GGUF deployment, the repository says Bonsai 2 requires PrismML's fork of llama.cpp; stock binaries cannot run these files. That compatibility requirement matters for developers evaluating local coding assistants, private document analysis or agents on existing hardware.