Qianwen AI Platform has launched Qwen3.8-Flash, a new model with a novel architecture that offers high performance at very low API prices, especially suited for agent tasks. The model weights are open-sourced on Hugging Face and ModelScope, laying the groundwork for Qwen4.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
The Qwen AI platform has released Qwen3.8-Flash, entering the high-frequency Agent invocation scenario at an input price of 0.8 yuan per million tokens, backed by efficiency gains from a new architecture. This article breaks down its technical highlights, pricing strategy, and open-source ecosystem, exploring whether this 'extreme cost-performance' model can define the next generation of large model competition.
On August 27, the Qwen AI platform officially launched Qwen3.8-Flash, touting 'extreme cost-effectiveness.' According to official announcements, its input price is as low as 0.8 yuan per million tokens, output at 2.7 yuan, and cache hits reduced to 0.1 yuan. This pricing is significantly lower than mainstream large models, making it especially suitable for scenarios like Agents that require frequent, high-volume calls.
The price advantage is not merely a subsidy but stems from cost reductions driven by architectural innovation. Officials state that Qwen3.8-Flash uses a multimodal mixture-of-experts (MoE) architecture, with 100 billion total parameters but only 6 billion activated, greatly reducing inference computation. Additionally, the self-developed QSA (Qwen Sparse Attention) mechanism, integrated with GDN, can speed up 1M long-context scenarios with high cache hit rates by over 8 times, directly lowering unit costs.
All information in this article comes from the official Alibaba Cloud Developer public account, which is first-party promotional material and has not been independently verified by third parties. All technical indicators and prices are as officially claimed; actual performance and cost-effectiveness should be based on real-world testing.
Primary report
Primary source