# Alibaba unveils 2.4-trillion-parameter model as DeepSeek cuts costs > Source: > Published: 2026-08-03 15:40:23+00:00 Alibaba unveiled Qwen3.8-Max on Monday, its largest artificial intelligence model to date, while DeepSeek released an updated model with lower API costs, underscoring the intensifying race among Chinese developers to compete on scale, performance and affordability. Alibaba Cloud said Qwen3.8-Max contains 2.4 trillion parameters and can process up to one million tokens within a single context window. The model is already available to developers through Alibaba Cloud Model Studio, with the full model set to be made available next week. **Alibaba scales Qwen3.8-Max to 2.4 trillion parameters** Rather than running its full 2.4-trillion-parameter network at once, Qwen3.8-Max selects a 95-billion-parameter subset for each request through a sparsely activated expert system. Alibaba said that design was intended to reduce computing demands and response times while supporting text, visual analysis, coding, research and tasks that require extended execution. The company also said the model could process lengthy documents and video content and turn the material into searchable knowledge bases. According to Alibaba, Qwen3.8-Max ranked fifth in Text Arena, second in Vision Arena and fourth in Frontend Code Arena, while an internal trial saw the model work on a software engineering project for 16 days. **DeepSeek pushes down V4-Flash API pricing** DeepSeek, meanwhile, said its DeepSeek-V4-Flash-0731 model entered public beta on July 31. The release retains the architecture and size of the earlier V4-Flash preview but underwent additional post-training aimed at improving coding and agent-based tasks. DeepSeek has made pricing a central part of the V4-Flash release, charging $0.14 per million uncached input tokens, $0.28 per million output tokens and $0.0028 per million cached input tokens. The model also supports reasoning and non-reasoning modes, a one-million-token context window and outputs of up to 384,000 tokens. **The U.S-China AI gap keeps shrinking** Moonshot AI’s [Kimi K3 adds to the pressure](https://thecoinheadlines.com/tech-and-ai/chinas-moonshot-ai-unveils-kimi-k3-closing-gap-with-u-s-ai-leaders/article-26572/) from Chinese developers seeking to narrow the gap with leading U.S. systems. Released in mid-July, the Beijing startup’s 2.8-trillion-parameter model combines native visual understanding with a one-million-token context window and is designed for extended coding, research and agent-based work. The new model selects 16 of 896 expert modules for each token to handle large workloads more efficiently, while Moonshot has also made the model files available to developers, unlike many proprietary systems. Together, these releases show China’s AI race moving beyond raw benchmark scores into a broader contest over scale, lower costs and developer access, a combination that could speed up adoption, intensify pressure on U.S. leaders and reshape how quickly advanced models reach businesses and users worldwide.