[Qwen3.8-Flash Open Source: 6B Activation, Multiple Agent Evaluations Surpass Opus4.6] Beating AI Newsflash: Alibaba's Qwen has released Qwen3.8-Flash and simultaneously open-sourced the model weights for Qwen3.8-Flash-Next. The open-source version includes an additional 'Next,' indicating it is the first to utilize the next-generation Qwen4 architecture. This is a multimodal MoE (Mixture of Experts) model, with the main model containing 125B parameters and activation parameters only 6B. Additionally, it incorporates a 51B N-gram Embedding, essentially a large-scale "lookup memory," trading more storage for reduced computation. In official evaluations, SWE-bench Pro scored 9.1 points higher than Claude Opus4.6, JobBench scored nearly 20 points higher, AndroidWorld scored 22.5 points higher, and MathVision scored 25.1 points higher. The model natively supports 262K context length, expandable to 1M. Pricing is also set very low. The production version has simultaneously launched Qwen Cloud, which by default supports 1M context length and tool invocation. The cost is 1 yuan per million tokens for input and 3 yuan for output. According to the official statement, Qwen3.8-Flash achieves capabilities close to Qwen3.7-Plus overall, but with training costs reduced to approximately 1/9. Qwen aims to deliver flagship-level model capabilities at a significantly lower computational cost. [Original Link]