Flow - 4x faster than typing
Mini toolsTernary Bonsai 27B
By PrismML
Model family: Qwen
Ternary Bonsai 27B is a ternary weight quantization of a 27B hybrid attention causal language model, using a 3 value weight alphabet with group wise FP16 scaling for roughly 1.71 bits per weight. It retains 80.49 average across 15 thinking mode benchmarks, close to the 85.07 full precision score, with math at 93.40, coding at 85.96, and agentic tool use at 74.01. It supports a 262K token context with 4 bit KV cache quantization, ships with a DSpark speculative decoding drafter for faster CUDA serving, and includes an optional vision tower for image input. Designed for laptop local agents, privacy sensitive offline use, and single GPU serving where full 27B class quality is needed within a small memory footprint.
New
Multimodal
Overview
A 27B parameter language model compressed to ternary (1.71 bit) weights, retaining 95% of full precision intelligence while shrinking deployed size to about 7.2 GB. Runs full 27B class reasoning, math, and coding on a standard laptop or single GPU with a 262K token context window.
About PrismML
Industry:
Artificial Intelligence
Company Size:
6
Location:
Pasadena, California, US
Website:
prismml.com
Last updated: July 15, 2026
