An updated 2.3bpw model for qwen 27b + 12 GB cards: ties Qwen3.6 BF16 on LiveCode Bench and reduces error by 37-57% vs a previous checkpoint
Hello everyone! A few weeks ago, I released a 2.3bpw model intended for the 30xx 12 GB Ampere cards. Using a custom version of sjkvarn and several runtime changes, this made it possible to hit 200K ctx on a single 12GB card. Benchmarking placed it somewhere around qwen3.5 27b, but below 3.6.