China’s artificial intelligence competition is widening from a race for the largest and most capable models into a battle over how cheaply companies can operate them.
Alibaba unveiled Qwen3.8-Max on August 3, describing the 2.4-trillion-parameter system as its largest and most capable AI model to date.
Alibaba Pushes Model Scale Beyond Two Trillion Parameters
Reuters said Qwen3.8-Max contains 2.4 trillion parameters, placing it close to Moonshot AI’s Kimi K3, which launched with 2.8 trillion parameters in July. Parameters are numerical settings learned during training that help a model recognize patterns, produce responses and complete tasks, although a larger total does not automatically guarantee better performance.
CTech reported that the model uses a mixture-of-experts architecture that activates only 95 billion parameters for an individual request. By directing work to specialized sections rather than running the entire system, Alibaba aims to reduce computing requirements and response times despite the model’s enormous overall size.
Qwen3.8-Max can process as many as one million tokens in a single context, allowing it to examine extensive codebases, lengthy legal files or hundreds of pages of documents. Alibaba also said the system completed a software-engineering project over 16 days and is scheduled for release the following week.
Early Rankings Lift Alibaba Shares
Alibaba’s Hong Kong-listed shares climbed 7% after Qwen3.8-Max appeared near the top of crowdsourced model rankings. On Arena.AI, the model became the highest-ranked Chinese system for text tasks, although several models from Anthropic remained ahead of it.
Qwen3.8-Max placed second worldwide on Arena.AI’s rankings for image analysis and visual reasoning, trailing only a version of Anthropic’s Claude Fable 5. The early results suggest Alibaba is trying to compete through both scale and multimodal performance rather than parameter count alone.
DeepSeek Competes Through Ultra-Low Pricing
DeepSeek’s V4-Flash is pursuing a different advantage: making artificial intelligence inexpensive enough for wider business and developer use.
Research firm Artificial Analysis ranked V4-Flash as the least expensive to operate among the well-known models included in its benchmark testing. DeepSeek charges $0.14 for every million input tokens and $0.28 for every million output tokens.
Artificial Analysis calculated an average benchmark cost of only $0.03 per test for V4-Flash. That compared with approximately $0.86 for Kimi K3, $1.86 for OpenAI’s GPT-5.6 Sol and $3.15 for Anthropic’s Claude Fable 5.
The pricing comparison considers the amount of information a model must process and produce to finish a task, not simply its advertised token rate. CTech explained that a model with a low headline price can still cost more when it needs additional computation or repeated steps to reach an answer.
Performance Still Separates the Cheapest Models
Lower operating costs do not necessarily mean that V4-Flash matches the strongest competing systems.
Artificial Analysis gave V4-Flash a score of 50 out of 100 on its Intelligence Index, matching Google’s Gemini 3.6 Flash but falling below Kimi K3’s score of 57. Anthropic’s Claude Opus 5 and Claude Fable 5, along with OpenAI’s GPT-5.6, scored at least nine points higher.
Reuters quoted Omdia chief analyst Lian Jye Su as saying many business workflows do not require the industry’s most advanced model and instead need systems that are affordable, accessible, transparent and sufficiently capable.
Alibaba and DeepSeek are therefore applying different pressure to the global AI market. Qwen3.8-Max challenges rivals with scale and strong benchmark performance, while V4-Flash tests how low prices can fall before cost becomes more important than having the most powerful model.
