Alibaba this week open-sourced the weights of Qwen3-Max's flagship (2.4 trillion parameters), runnable on NVIDIA's latest GB300 GPUs. What deserves attention: it's one of the largest models publicly downloadable today, with claimed performance approaching the closed-source frontier. Chinese companies now have, for the first time, a real option to access near-top-tier AI without being locked into OpenAI or Anthropic.

What this is

Qwen3-Max is the latest flagship of Alibaba's Tongyi Qianwen series. This open-weight release is the 2.4 trillion parameter variant (2.4T-A95B), which activates only 95 billion parameters per inference. This "on-demand expert routing" design is called MoE (Mixture of Experts)—total parameters are massive, but per-query cost stays controllable.

Key points:

  • Hybrid attention: combines full attention (high precision) with linear attention (cheap for long text), reportedly keeping costs manageable even at 1M-token context windows
  • Configurable inference: users can toggle between "deep thinking" and "fast response" modes, trading cost against latency depending on the task
  • Deployment barrier: requires data-center-grade hardware like NVIDIA GB300 NVL72; a single deployment starts at tens of millions of yuan (low eight figures USD)

Industry view

Positive voices come mainly from Alibaba Cloud customers and the domestic AI community: open weights give heavily regulated industries—finance, government, energy—where "data can't leave the premises" their first real shot at self-hosted large models. NVIDIA is clearly pleased—each flagship model tuned for its hardware means another batch of orders.

But skepticism is plentiful:

  • For most companies, true "open source" is an illusion—the 2.4T model simply cannot run without hundreds of millions of yuan in hardware budget
  • Alibaba simultaneously sells API access, compute, and consulting services; the tension between open weights and commercial interests raises doubts about long-term community maintenance
  • A big model doesn't mean strong in every scenario; enterprises still need extensive fine-tuning and engineering, and post-deployment performance may disappoint
  • Overseas analysts are more measured: open-source catching up with closed-source relies mainly on parameter scaling; gaps remain in alignment, safety, tool use, and other engineering details

Impact on regular people

For enterprise IT: heavily regulated industries—finance, government—can now seriously evaluate the cost structure of "self-hosted large models" rather than being forced down the paid-API-only path.

For working professionals: your daily ChatGPT or Claude experience won't change in the short term. But if your company's backend swaps to Qwen3-Max, a "silent upgrade" to conversation quality may show up in your DingTalk and enterprise knowledge base within one to two months.

For consumer markets: AI assistants across Alibaba's ecosystem—DingTalk, Taobao, Amap—will almost certainly show visible quality leaps within the year, since they're now backed by this new model.