Alibaba open-sources Qwen3.8-Flash with next-gen architecture, training cost at one-ninth of predecessor

On the evening of August 26, Alibaba released and open-sourced its latest Qwen model Qwen3.8-Flash. The model uses a new next-generation (Next) architecture, with 125B Transformer parameters and only 6B active, achieving frontier performance that surpasses Claude Opus 4.6 and approaches Opus 4.8.

Thanks to comprehensive architectural and training innovations, Qwen3.8-Flash's training cost dropped nearly 90% compared to Qwen3.7-Plus, with inference cost also significantly reduced. Pricing is 1 yuan input and 3 yuan output per million tokens, as low as one-third of DeepSeek-V4-Flash. Weights have been open-sourced on Hugging Face and ModelScope, with FP8 quantized versions released simultaneously.

In terms of architecture, the model uses a unique QSA (Qwen Sparse Attention) mechanism in the attention component, combined with GDN to form an efficient hybrid attention architecture, achieving over 8x speedup in high-cache-hit 1M token long-context scenarios. The residual component introduces Gated Residual, splitting the single traditional Transformer information channel into four parallel channels with dynamic gates. The embedding component introduces 51B N-gram embedding parameters that can be offloaded to host memory. The optimization stage uses the Muon Optimizer with a refitted Scaling Law.

On SWE-bench Pro, Qwen3.8-Flash leads Opus 4.6 by 9.1 points. On CoWorkBench and Toolathlon Verified agent tasks, it surpasses DeepSeek-V4-Flash. On JobBench, it exceeds Opus 4.6 by nearly 20 points. In multimodal capability, it leads Opus 4.6 by 22.5 points on AndroidWorld, 25.1 points on MathVision, and 31.5 points on embodied intelligence ERQA tasks.

The new Next architecture is officially considered the prototype for the next-generation Qwen4 models. The Qwen3.8 series has open-sourced three sizes: the 2.4T parameter Qwen3.8-Max, Qwen3.8-27B, and Qwen3.8-Flash.