# Ternary Bonsai 2 27B

PrismML, released 2026-09-17

**Verdict: conditional.** Ternary Bonsai 2 27B squeezes Qwen3.8 27B into about 5.9GB at 1.76 effective bits, so a 262K-context text-and-image model fits on a Mac or a single consumer GPU under Apache 2.0. Conditional because the quality figures are PrismML's own, its blog and model card report different aggregate scores, and the files need PrismML's custom kernels: stock llama.cpp will not run them. Right for local-inference users short on memory. Last verified 2026-10-02.

## Pricing

- Input: Free (open weights)
- Output: Free (open weights)
- Context: 262K ctx; self-hosted; PrismML kernels required

## Strengths

- Running a 27B-class model in about 6GB: the ternary GGUF is 5.95GB against a 53.8GB FP16 reference, plus a 0.63GB projector for image input
- Apple silicon Macs: PrismML reports 46.8 tokens per second on an M5 Max through MLX
- Long-context local work with a 262K-token window
- Commercial use and redistribution under Apache 2.0
- Standard local tooling: the GGUF files need PrismML's llama.cpp fork, and stock llama.cpp will not run them
- Maximum quality: PrismML's own figures put it at 98.2% of full-precision Qwen3.8 27B
- Decisions that need independent evaluations: all published scores are PrismML's

## Benchmarks

- Aggregate retention vs Qwen3.8 27B full precision (PrismML, thinking mode): 98.2% (https://prismml.com/news/bonsai-2-27b)
- Math category (PrismML, thinking mode): 96.57 (https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf)
- Decode speed, Apple M5 Max (PrismML): 46.8 tok/s (https://prismml.com/news/bonsai-2-27b)

## Sources

- [Hugging Face: PrismML Bonsai 2 collection](https://huggingface.co/collections/prism-ml/bonsai-2)
- [PrismML: Bonsai 2 27B announcement](https://prismml.com/news/bonsai-2-27b)
- [Hugging Face: prism-ml/Ternary-Bonsai-2-27B-gguf model card (also PrismML announcement)](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf)

Canonical: https://neomanex.com/models/ternary-bonsai-2-27b