Unsloth updates compressed Qwen3.8-27B models, claiming accuracy gains

Unsloth released updated compressed versions of Qwen3.8-27B on November 11, saying its Dynamic V3.1 preview algorithm improves average accuracy by 5-10% at comparable file sizes. Its tests include Terminal Bench 2.1 and DeepSWE v1.1 for agent tasks, alongside coding and general benchmarks. The 27B-parameter model supports vision, reasoning and a 256K context window, and can run locally through llama.cpp and Unsloth Desktop. Unsloth says four-bit versions work with 16-19GB of VRAM. Most versions from two-bit through six-bit were updated; three-bit UD3 versions remain unchanged because the company found their performance comparable to UD3.1.