China's AI competition intensifies as Alibaba unveils its biggest model and DeepSeek drives costs lower - calcalistech.com

China's AI competition intensifies as Alibaba unveils its biggest model and DeepSeek drives costs lower

Alibaba is betting on scale with a 2.4-trillion-parameter model, while DeepSeek is focusing on dramatically lower costs as Chinese companies intensify competition with U.S. AI leaders.

China's artificial intelligence race entered a new phase on Monday as Alibaba unveiled its largest AI model to date, just days after DeepSeek released what independent researchers say is the cheapest major AI model in the world to operate.

Together, the announcements underscore how competition among Chinese AI companies has evolved beyond simply building more capable systems. The industry's new battleground is balancing performance, scale and cost, as companies compete to win developers and enterprise customers while challenging U.S. AI leaders.

Alibaba introduced Qwen3.8-Max, describing it as its most capable model yet. With 2.4 trillion parameters, it is only slightly smaller than rival Moonshot AI's recently launched Kimi K3, which contains 2.8 trillion parameters.

Although parameter count does not necessarily determine model quality, it has become one of the most closely watched indicators of the scale of modern AI systems. Unlike U.S. companies such as OpenAI, Anthropic and Google, which keep details of their flagship models private, Chinese developers have increasingly embraced open-weight models, allowing developers to download and adapt the underlying model weights. Publishing parameter counts has become another way for companies to demonstrate technological progress and attract developers.

Alibaba said Qwen3.8-Max uses a mixture-of-experts architecture, meaning only 95 billion parameters are activated for any individual request despite the model's overall size. By routing tasks to specialized parts of the model instead of activating the entire system, the company says it can reduce both computing costs and response times.

The model can process up to one million tokens at a time, allowing it to analyze extremely large inputs, including lengthy legal documents, extensive software codebases or hundreds of pages of text. Alibaba also said the model recently completed a software engineering project in 16 days. Qwen3.8-Max is scheduled to become available next week through Alibaba Cloud's Model Studio platform.

The launch comes as Alibaba faces increasingly intense domestic competition.

DeepSeek, whose low-cost AI strategy shocked global markets last year, released its V4-Flash model on Friday, reinforcing its position as the industry's cost leader.

According to research firm Artificial Analysis, V4-Flash costs just $0.03 per benchmark test on average, far below competing systems. The firm estimates Moonshot's Kimi K3 costs approximately $0.86 per test, OpenAI's GPT-5.6 Sol around $1.86, while Anthropic's Claude Fable 5 reaches $3.15.

Artificial Analysis argues that benchmark costs provide a more meaningful comparison than headline pricing because they account for how much data a model must process and generate to complete a task. A model with lower pricing per token may ultimately prove more expensive if it requires substantially more computation to reach an answer.

DeepSeek's published pricing also reflects its aggressive strategy. The company charges $0.14 per million input tokens and $0.28 per million output tokens, making it substantially cheaper than many competing models.

The startup first drew global attention in early 2025, when its R1 model triggered a broad selloff in technology stocks by raising questions about whether leading U.S. AI companies were overspending on increasingly expensive infrastructure.

That early advantage, however, has quickly given way to fierce competition inside China itself.

DeepSeek now faces challenges not only from startups including Moonshot AI, MiniMax and Z.AI, but also from technology giants such as Alibaba and ByteDance, all competing to attract businesses looking for affordable AI systems that can be deployed at scale.

Performance remains another area of competition.

Artificial Analysis assigned V4-Flash a score of 50 out of 100 on its Intelligence Index, matching Google's Gemini 3.6 Flash and trailing Moonshot's Kimi K3, which scored 57. Anthropic's Claude Opus 5, Claude Fable 5 and OpenAI's GPT-5.6 scored at least nine points higher.

Alibaba's new model also posted strong early results. On Arena.AI's crowdsourced rankings, Qwen3.8-Max immediately became the highest-ranked Chinese model for text generation, although it still trails several Anthropic models overall. In multimodal tasks involving images and visual reasoning, it ranked second globally behind a Claude Fable 5 variant.

The competition is unlikely to slow.

Reuters reported that DeepSeek is already preparing a more powerful successor, V4-Pro, although the company has not announced a launch date.