Alibaba Launches Qwen3.8-Flash AI Model with 89% Lower Training Costs, Eyes Qwen4 Prototype

Alibaba's Qwen launches Qwen3.8-Flash AI model with lower training costs

Qwen3.8-Flash supports ⁠a default context window of 262,144 tokens, expandable to 1 million tokens, enabling it ⁠to handle large files, lengthy conversations and extensive research materials. It also released open-source weights for Qwen3.8-Flash-Next, allowing ⁠the developer community to evaluate an architecture that Qwen said would serve as a prototype for its next-generation Qwen4 model family. Alibaba's Qwen has released a new multimodal model, Qwen3.8-Flash, that delivers stronger coding and office-task performance while cutting training costs, it said in a statement on Wednesday published on social media. As ecommerce growth stagnates, ⁠AI ⁠has become the biggest driver of revenue growth for the Chinese e-commerce and cloud computing giant. Its Qwen AI models are some of the most popular in China, where competition is intensifying in the sector. Qwen3.8-Flash supports ⁠a default context window of 262,144 tokens, expandable to 1 million tokens, enabling it ⁠to handle large files, lengthy conversations and extensive research materials. Compared with its previous Qwen3.7-Plus model, Alibaba said it requires about one-ninth of the training cost. Qwen said it would charge 1 yuan ($0.1488) per million input tokens and 3 yuan per million output tokens for access through its application programming interface. It also released open-source weights for Qwen3.8-Flash-Next, allowing ⁠the developer community to evaluate an architecture that Qwen said would serve as a prototype for its next-generation Qwen4 model family. On Sunday, Alibaba launched a heavily discounted HK$80-billion share sale to fund its AI ambitions. ($1 = 6.7220 Chinese yuan renminbi)

Alibaba Targets Coding and Office Tasks With Low-Cost AI Model | PYMNTS.com

Qwen3.8-Flash is a multimodal mixture of experts (MoE), and the production version will be available soon via the QwenCloud API at 16 cents per 1 million input tokens and 47 cents per 1 million output tokens. It features 125 billion parameters and 51 billion N-gram embeddings, with 6 billion activated per token, according to the post. "Dramatically lower training and inference costs: trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board with especially strong gains in coding and office tasks," Alibaba said in the post. It was reported Aug. 15 that Alibaba's open-weight AI models have become the world's top AI offerings. The company's models have seen more than 3 billion downloads worldwide over the last six months, surpassing those of Meta, Google and DeepSeek. Alibaba said Qwen, its family of AI models, has open-sourced more than 460 models and its ecosystem has issued over 300,000 derivatives. Findings from open-source AI hub Hugging Face showed that Google had 418 million downloads for the year, and Meta had 227 million. It was reported Aug. 14 that Apple worked with Alibaba Group to develop and train an Apple AI model specifically for the Chinese market. Apple has previously used third-party models to power the AI features it offers in the iPhones and other devices it sells in China. Now, following the collaboration with Alibaba, the company plans to launch its Apple Intelligence suite of AI tools in China within months. Having its own model tailored for China would give Apple greater control over the AI experience it offers in the market and would make Apple the first foreign company to secure Chinese government approval to offer a proprietary model in the country. Alibaba reported Thursday (Aug. 20) that during the most recent quarter, its AI-related product revenue saw triple-digit growth for the 12th consecutive quarter. "With our full-stack AI strategy, we have put Alibaba in a superior position to capture the substantial growth of demand for artificial intelligence and AI compute," Alibaba CEO Eddie Wu said in an earnings press release. For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.