Qwen3.8-Flash Released with 6B Activations and Outperforms Opus4.6 in Agent Evaluations

Dongcha Beating AI News: Alibaba's Qwen has released Qwen3.8-Flash and simultaneously open-sourced the model weights as Qwen3.8-Flash-Next. The open-source version includes an additional "Next," indicating it is the first to adopt the next-generation Qwen4 architecture. This is a multimodal MoE model with a main model of 125B parameters, but only 6B activated parameters. It also adds 51B N-gram embeddings, effectively a large-scale "lookup memory" that trades increased storage for reduced computation. In official benchmarks, Qwen3.8-Flash outperforms Claude Opus4.6 by 9.1 points on SWE-bench Pro, nearly 20 points on JobBench, 22.5 points on AndroidWorld, and 25.1 points on MathVision. The model natively supports 262K context length and can be extended to 1M. Pricing is also significantly reduced. The production version simultaneously launches Qwen Cloud, which defaults to 1M context length and tool calling, at a cost of 1 yuan per million input tokens and 3 yuan per million output tokens. According to the official team, Qwen3.8-Flash achieves near the overall capability of Qwen3.7-Plus while reducing training costs to approximately one-ninth. Qwen aims to deliver near-flagship model performance at a much lower cost by using less computation.