Megadose AI progress, ranked and analyzed.

PrismML releases Bonsai 2 27B, which compresses Alibaba's Qwen3.8 27B to 5.9 GB, small enough for smartphones, while retaining 98.2% of Qwen's benchmark scores (Julie Bort/TechCrunch)

Techmeme ·
PrismML says Bonsai 2 27B keeps almost all of Qwen3.8 27B’s measured performance in a 5.9 GB model.

Techmeme’s note cites TechCrunch’s Julie Bort and PrismML’s release of Ternary Bonsai 2 27B. The model is based on Alibaba’s Qwen3.8 27B and is described as 9x smaller than the full-precision version. PrismML says it retains 98.2% aggregate performance across 20 benchmarks, with the footprint still small enough for local use on phones or ordinary laptops. Techmeme’s note

score 6

Categories: Model Releases