PrismML's Ternary Bonsai 2 squeezes Qwen3.5 VL models to ternary weights, 9B at 2GB
lmoroney · x · 2026-10-20
PrismML released Ternary Bonsai 2, converting Qwen3.5 vision-language models at 2B/4B/9B into ternary weights (-1/0/+1), with files from 0.43 GB to 2 GB. The company claims the 9B keeps over 90% quality across 19 benchmarks, and weights are Apache 2.0. The author explains how ternary weights work, walks the benchmark table row by row, and lists what to verify before putting one on a phone.
Related event: PrismML Shrinks Qwen3.5 VL to Ternary Weights, 9B Model Fits in 2GB(2 posts)→
More from Models
- Big-Big-SycoBench: GPT-4o caves to invalid pushback 94.6% of the time, Opus 5.5 accepts valid criticism only 32.2% — i_dg23 · 2026-10-20
- OpenAI said to be reviving Terra-class models to fill its cheap-fast tier gap — haider1 · 2026-10-20
- PlotGround benchmark: best multimodal model hits 87.5% at ±5% reading scientific plots — james_y_zou · 2026-10-20
- MiniMax confirms Space Bunny was early M3.1 Flash: $0.15/M input, open weights Oct 23 — lmoroney · 2026-10-20
- Meta Muse hits 5M downloads in 22 days, beating ChatGPT's pace — aakashgupta · 2026-10-20
- BioEmu2 tease: successor to the protein dynamics model may drop soon — sokrypton · 2026-10-20