PrismML releases Ternary Bonsai 2 27B: under 6GB with 98.2% of FP16 performance

On September 18, PrismML released Ternary Bonsai 2 (27B) on Hugging Face, a ternary-quantized inference model based on Qwen3.8-27B. The model quantizes weights to the ternary values -1/0/1, compressing its size to 5.95GB (versus 54GB for the FP16 original, roughly a 9x reduction) while retaining 98.2% of aggregate benchmark performance, and can run directly in a WebGPU browser.

Confirmed

Why it matters

Ternary quantization compresses weights to -1/0/1, dramatically reducing memory footprint and inference cost, and is a representative technique for on-device and low-cost deployment. Multiple sharers noted that a 6GB-class size lets a 27B-parameter model run on consumer devices or even in a browser—if the quality loss is controllable, it could change how large models are distributed and used.

2026-09-18 ~ 2026-09-18 · 6 related posts

Primary sources

1 near-duplicate retellings: cephaloform