Unsloth releases Qwen3.8-Flash-Next GGUFs, claims 10-15% accuracy gain

Unsloth AI releases Qwen3.8-Flash-Next GGUF quantizations and says they score 10-15% higher in accuracy than GSQ-RCO on DeepSWE, Terminal Bench and Aider Polyglot.

The company credits its Unsloth Dynamic v3.1 quantization method and says the models run on setups with 32GB or more of RAM.

An earlier Unsloth post on X gave the gain as 15% and said Dynamic v3.1 beats other quantizations by more than 15% at a 256K context. A follow-up post from Unsloth's Daniel Han narrowed the range to 10-15% and named GSQ-RCO as the baseline. Neither post gives per-benchmark scores.

Unsloth also published root mean square error against BF16 for three formats: Q8_0 at 54.4 GB (0.0053), Q5_1 at 38.4 GB (0.0376) and IQ4_NL at 28.8 GB (0.0757).