Qwen3.8-Flash-Next can run locally with Unsloth

Qwen3.8-Flash-Next is a new open-weight, 125B parameter MoE multimodal model from Qwen built on the Qwen4 architecture. It supports a 262K context window, includes advanced reasoning features and is presented as outperforming Claude-4.6-Opus (Max) while running locally on devices with 75GB RAM or unified memory and no GPU VRAM required.

Unsloth provides GGUF builds for running the model through Unsloth Desktop or llama.cpp. The smallest 1-bit quantized version uses 4-bit for the Ngram or PLE components, is 79% smaller than BF16 (355GB), and retains a top-1 accuracy of 80%. Hardware guidance lists memory needs ranging from 75 GB for 1-bit to 355 GB for BF16, with 96GB RAM or unified memory recommended for the smallest configuration.

The model supports thinking and non-thinking modes with separate default sampling settings, plus Preserve Thinking for continued conversations. MTP is enabled by default in Unsloth Desktop and can deliver 1.3 to 1.7x faster inference, including 170 tokens/s on 1x RTX 6000 PRO GPU compared with a 100 token baseline.