Bonsai 2 27B vs Qwen 3.8 27B

Analysis

Bonsai 2 27B is a proprietary PrismML requantization of Qwen 3.8 27B, not a separate model with its own weights: same architecture, same training, same 27B parameters, run through a 1.58-bit (ternary) compression pass. That makes this comparison unusual for this site: there's no capability tradeoff to discover, only a size-versus-quality one, since every difference in the numbers below comes from compression, not a different pretrain.

Qwen 3.8 27B ships the standard FP32-to-Q2_K ladder: BF16 weights run 54.0 GB, and Alibaba's own recommended Q4_K_M build totals 19.0 GB at 8k context (16.4 GB weights, 0.5 GB KV cache, 12% activation overhead). Bonsai 2 27B skips that ladder entirely and ships one fixed Ternary (Q2_0) file: 5.9 GB weights, 7.2 GB total at 8k context, roughly 9.2x smaller than Qwen 3.8 27B's own BF16 weights and still 2.6x smaller than its recommended Q4_K_M build. PrismML's own six-category capability table, run against this exact FP16 checkpoint, reports 98.2% aggregate quality retained at that footprint, with the smallest hit on Vision (96.3% retention) and, unusually, a small gain on Instruction Following (82.66 vs 81.25, a 102% retention figure). That capacity difference isn't just about card size: because Qwen 3.8 27B's real hybrid stack (16 of 64 layers keep a KV cache) carries over into Bonsai 2 27B unchanged, the KV-cache math is identical between the two; the entire VRAM gap comes from the weights file alone. It also shows up in GPU reach: 111 of the 119 GPUs this site tracks run Bonsai 2 27B natively at 8k context against 84 for Qwen 3.8 27B, and every GPU that fits Qwen 3.8 27B also fits Bonsai 2 27B, plus 27 more, mostly older or 8-12GB consumer cards that can't clear even Qwen 3.8 27B's lowest real ladder rung.

Bottom line: If an 8GB-class GPU, base-tier Apple Silicon, or a phone is the hard constraint, Bonsai 2 27B is the only one of the two that actually runs there, and PrismML's own numbers say it keeps most of the original's quality intact. If the hardware isn't the constraint, Qwen 3.8 27B itself is the safer choice: 1.58 bits/weight is still an aggressive compression, PrismML's retention figures are vendor-reported and not yet independently verified, and a standard Q4_K_M build already fits comfortably on any 24GB card with far less quality risk. Reach for the Ternary build specifically because the VRAM ceiling forces it, not as a free way to get the same model for less.