Qwen 3.8 Flash Cheaper Than DeepSeek in Real-World Test, Disrupting the Market

Alibaba has open-sourced Qwen 3.8 Flash-Next, featuring a 125B MoE architecture (with ~6B activated parameters), supporting a 262,144-token context window. Training costs have dropped significantly, performance has improved, and pricing is only 1/3 that of DeepSeek-V4-Flash. Real-world tests show it can quickly handle office tasks such as copy rewriting and meeting minutes summarization. Meanwhile, Zhipu has also open-sourced GLM-5.3-Flash, with performance comparable to Claude Opus 4.8 but at a lower price, as the domestic large-model price war among Chinese players continues to intensify.