Bonsai 2 27B vs Bonsai 27B

Analysis

Bonsai 2 27B and the original Bonsai 27B are both proprietary low-bit requantizations from PrismML, same idea, different base checkpoint. The original (July 2026) requantizes Qwen 3.6 27B; the sequel (September 2026) requantizes the newer, stronger Qwen 3.8 27B. Since both base models share the same real hybrid architecture (16 of 64 layers keep a KV cache; the other 48 run linear attention with a fixed-size state), the KV-cache math is identical between the two Bonsai releases; the interesting differences are in what PrismML actually shipped and how much quality each retained.

The original Bonsai 27B ships two fixed builds: a 1-bit (Q1_0) file at 3.8 GB weights retaining 89.5% of FP16 quality, and a Ternary (Q2_0) file at 7.2 GB retaining 94.6%. Bonsai 2 27B ships only one: a Ternary (Q2_0) file at 5.9 GB, smaller than the original's own Ternary build despite the identical 27B parameter count and the same 1.58 bits/weight, so PrismML's packing improved on top of retargeting the base model. It also retains more: PrismML's own six-category capability table for Bonsai 2 27B reports 98.2% aggregate retention against Qwen 3.8 27B, against the original's 94.6% retention against Qwen 3.6 27B, and Bonsai 2 27B is the rare case where one category (Instruction Following) actually scores above its FP16 base rather than below it. Neither release has independent third-party benchmark verification yet; both are PrismML's own reported figures.

Bottom line: For anyone choosing between the two today, Bonsai 2 27B is the better pick: a smaller download than the original's Ternary build, higher reported quality retention, and a newer, stronger base model underneath. The only reason to reach for the original instead is its 1-bit build, at 3.8 GB the smallest Bonsai file that exists; Bonsai 2 27B doesn't offer an equivalent sub-4GB option, so a 6-8GB device that can't fit even 5.9 GB is the one case where the original's extra build still matters.