Alibaba Says Its New Model Is Second Best In The World

Alibaba previewed Qwen 3.8-Max at the World Artificial Intelligence Conference in Shanghai and made a specific claim: the 2.4-trillion-parameter model trails only Anthropic's Claude Fable 5. No benchmark data came with it. No model card. No confirmation of how many of those 2.4 trillion parameters are actually active during inference. What did come with it was OpenAI and Anthropic API compatibility and a 10% promotional discount, aimed squarely at developers already comfortable switching providers.

Key Points:

• Alibaba previewed Qwen 3.8-Max, a 2.4-trillion-parameter multimodal model, claiming it's second only to Claude Fable 5, without publishing benchmarks, a model card, or active parameter counts.
• The launch directly targets Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model whose delayed release strategy has been fueling paid API adoption ahead of a planned IPO.
• Chinese-origin models now account for as much as 46% of weekly enterprise token volume on OpenRouter, up from an 11% average over the prior twelve months, largely because they run 60% to 90% cheaper than leading US models.
• Two House committees are jointly investigating US companies like Airbnb and Cursor-maker Anysphere over their use of Chinese-built models, citing data exposure and distillation concerns.
• Even among businesses actively using open-source or Chinese model platforms, more than 90% still also use OpenAI or Anthropic, suggesting task-based routing rather than wholesale replacement.

An Unverified Claim, Aimed At A Specific Rival

Alibaba's "second only to Fable 5" framing is a marketing position dressed as a technical one. Without published benchmarks or a model card, there's no way to independently confirm where Qwen 3.8 actually lands, and industry watchers see the timing as a direct shot at Moonshot AI's Kimi K3, whose delayed open-weight release has been driving paid subscriptions ahead of an IPO. By promising to eventually open-source Qwen's weights, Alibaba is trying to pull developers away from Kimi's paid tier before Moonshot converts that momentum into a public offering.

The Adoption Numbers Are Not Subtle

Whatever Qwen 3.8 turns out to be, it's launching into a market that's already moved. Chinese-origin models have held above 30% of weekly enterprise token volume on OpenRouter since early February 2026, and that figure has peaked as high as 46%, compared to an 11% average over the prior year and just 4.5% in the first half of 2025. The reason is almost entirely cost. DeepSeek's cheapest model runs at a fraction of a cent per million input tokens against roughly five dollars for a leading US model, and platforms like Z.ai's GLM 5.2 have posted adoption curves, 27 times the daily token volume and 80 times the customer count in a single week, that outpace anything US labs have shown this year.

Security Versus Convenience Versus Cost, In Practice

The Trump administration and two House committees aren't treating this as a neutral cost decision. The State Department has said Chinese models raise concerns about advancing state narratives and reflecting government-approved messaging, and a joint congressional investigation into Airbnb and Cursor-maker Anysphere is examining whether Chinese-built models expose US company data to foreign legal jurisdiction. At the same time, the economics are real enough that even risk-averse enterprises are testing the waters carefully: one data provider found that businesses using open-source or Chinese model platforms are still using OpenAI or Anthropic in the high 80s to mid 90s percent of cases, which points to task-based routing rather than a full switch. Sensitive workloads stay on trusted US models. Lower-stakes, high-volume work increasingly goes wherever it's cheapest.

What Business And Marketing Leaders Should Actually Weigh

This isn't a decision to make once and forget. Match the model to the task, not the vendor to your whole stack: a Chinese open-weight model handling internal drafts or bulk content generation carries a different risk profile than one processing customer data or proprietary strategy. Document your reasoning now, since regulatory scrutiny is already here and procurement decisions made quietly this year could face questions next year. And treat unverified benchmark claims, from any vendor, US or Chinese, as marketing until proven otherwise. Building that kind of vendor evaluation discipline into your operations is exactly the kind of work we fold into every growth strategy engagement.

If your team needs a clear-eyed framework for where Chinese models make sense and where they genuinely don't, our AI marketing services team can help you build one.

Source: South China Morning Post, "Alibaba says newest Qwen AI model is second only to Anthropic's Claude Fable 5"