Alibaba and DeepSeek Push Scale and Cost Boundaries

Alibaba unveiled its largest artificial intelligence model to date on 3rd August, while DeepSeek released a system that costs less than any well-known rival to run, two developments that highlight the diverging strategies Chinese firms are using to compete with U.S. AI leaders.

Alibaba's Qwen3.8-Max carries 2.4 trillion parameters, putting it close to the 2.8 trillion parameters of domestic rival Moonshot's Kimi K3. The model shot to the top of crowdsourced leaderboard Arena.ai as the highest-ranking Chinese text model, though it still trails Anthropic's Claude Fable 5 and several Opus variants. On visual tasks it placed second globally, behind only one Fable 5 variant. Alibaba shares jumped 7 percent in Hong Kong trading.

The model uses a mixture-of-experts design that activates only 95 billion parameters at a time, reducing computing costs and response delays. It can handle text, images, and video, and processes up to one million tokens at once. Alibaba said it completed a software-engineering project in 16 days and will formally release the model next week.

DeepSeek took the opposite approach. Its V4-Flash model, released on August 1, charges $0.14 per million input tokens and $0.28 per million output tokens. Research firm Artificial Analysis estimated its average cost at 3 cents per benchmark test, compared with 86 cents for Kimi K3, $1.86 for OpenAI's GPT-5.6 Sol, and $3.15 for Claude Fable 5. That makes V4-Flash more than 100 times cheaper than Anthropic's top model on a per-task basis.

Both models are open-weight, meaning developers can download and adapt the underlying settings. Omdia analyst Lian Jye Su said Chinese AI companies have found a market in business workflows that need models that are good enough, affordable, and transparent rather than the most powerful available. The two releases show Chinese firms attacking the global AI market from opposite ends: raw scale on one side, and cut-throat pricing on the other.