"Qwen 3.8-Flash"... Launch of the Cheapest AI Model
Alibaba Group announced the launch of the latest versions of its Qwen AI model series, a move aimed at reducing costs and expanding the global reach of its core platform. The Chinese company made the Qwen 3.8-Flash model available for download and revealed its weights, which are the values that help AI systems make decisions.
The new model comprises 125 billion parameters, along with 51 billion advanced linguistic embeddings, with only 6 billion actually activated when processing each token. According to the group’s statement, the new model’s performance rivals that of the latest releases from competitors, including Anthropic’s Opus 4.6 and DeepSeek’s V4-Flash.
The company clarified that the new version relies on an architecture originally designed for the next generation of models within the Qwen4 family, significantly lowering training and inference costs. Reuters reported that training Qwen 3.8-Flash required approximately one-ninth the cost of training its predecessor, Qwen 3.7-Plus, while demonstrating clear superiority in programming and office tasks.
The model supports a virtual context window of 262,144 tokens, expandable to one million tokens via YaRN technology, enabling it to process large files, extended conversations, and extensive research materials. Qwen set the API usage price at one yuan per million input tokens and three yuan per million output tokens, aiming to establish a new pricing standard in the market.
In this context, the company also released the open-source weights of the Qwen 3.8-Flash-Next model as a prototype paving the way for the upcoming Qwen4 architecture, allowing the developer community to test and adapt to the new structure. It achieved 84.5 points on AndroidWorld and 95.7 points on MathVision within specified confidence margins.
This move comes as AI work becomes the largest driver of Alibaba’s revenue growth, amid slowing growth in its e-commerce sector and intense competition from Chinese rivals such as ByteDance and DeepSeek.
