PrismML compresses Qwen3.8 27B to 5.93 GB with ternary weights

PrismML's Ternary Bonsai 2 27B takes Alibaba's Qwen3.8 27B and squeezes it from 53.80 GB down to 5.93 GB by holding every language-model weight to one of three values: minus one, zero, or plus one. According to MarkTechPost, the ternary scheme lands the model at roughly 1.72 bits per weight, with only 26.2 million parameters, or 0.0976% of the total, kept in higher precision for recurrent state paths and normalization. "Each group of 128 weights shares 1 FP16 scale," the writeup notes, and a blockwise Hadamard rotation inspired by SpinQuant is applied before ternary assignment.

On 20 benchmarks the compressed model averages 83.9 against the parent's 85.4, or 98.2% retention. Math holds up best at 99.5% and coding at 99.3%.

The weak spot is agentic work. On Terminal-Bench 2.1, Ternary Bonsai 2 scores 52.8 against the full-precision Qwen3.8 27B's 69.7, a gap of nearly seventeen points that the retention headline glosses over.

Throughput is the other selling point: 142.5 tokens per second on an RTX 5090 at 0.582 mWh per token, 46.8 on an M5 Max, 27.7 on an M5 Pro. The Apache 2.0 weights sit on Hugging Face, with MLX packs for Apple Silicon and a WebGPU browser demo, though loading them requires "PrismML's llama.cpp fork" because stock llama.cpp cannot parse the PTQ1_0 and PQ2_0 formats. It arrives amid a busy week of open-weight releases we've been logging in open-source AI.