ran
@PrismML's Bonsai 2 27B through my own agentic coding benchmark -- 20 multi-step bug fixes on
@pidotdev, ~55 turns each, all local on a single 3060.
Bonsai 2 27B (7.2GB): 13/20
GSQ-RCO 2-bit (8.4GB): 16/20
ByteShape 2-bit (8.8GB): 10/20
AtomicChat 2-bit (9.0GB): 7/20
Official Qwen API: 16/20
GSQ-RCO still holds the top spot. but bonsai beat every other local 2-bit quant I've run -- bytshape 10/20 at 8.8GB, atomicchat 7/20 at 9.0GB -- while being the smallest file of the three. at that point there's no reason to pick either of them.
and the PTQ1_0 pack of the same model is 5.9GB, so this lands in 8GB-card territory.