Alibaba's Qwen 4 API: 1.2T-parameter MoE at $1/$4 per million tokens, open weights promised by Dec 20

Alibaba has opened API access to Qwen 4, the first model of its next generation. It has 1.2 trillion parameters, with 56 billion active per token in a sparse Mixture-of-Experts design. It supports a 1M-token context window, accepts multimodal input and offers switchable hybrid thinking, so developers can choose between quick answers and extended reasoning. Pricing is $1 per million input tokens and $4 per million output tokens. Alibaba promises Apache 2.0 weights by December 20.

In parallel, Alibaba Cloud opened a two-week developer preview of Qwen4-Max, the flagship of the family, running through November 25 via Model Studio's Token Plan. The cloud-only Max model handles text, images and video, supports up to 393,216 output tokens, offers native function calling and is compatible with OpenAI and Anthropic API formats, so existing code can switch with minimal changes. Alibaba is focusing it on complex coding, front-end design and multi-step agentic office tasks, with temporary prices as low as 10% of list.

The pricing is the competitive weapon. At $1/$4, Qwen 4 costs a small fraction of Anthropic's listed $8/$40 for Fable 5.2. Community reaction has been positive, and the model reportedly ranked third on performance indices at launch. Qwen also has a strategic distribution win: Apple Intelligence's China rollout in December uses Qwen, after Apple cleared registration in Shanghai for iPad and Mac.

Caveats: the 'by December 20' weights promise is not yet fulfilled, and earlier Qwen flagships have sometimes stayed closed while smaller models were opened. Self-reported rankings need independent confirmation, and the low preview pricing is temporary. Some enterprises outside China will avoid a Chinese provider's hosted API regardless of price, which makes the open weights the key deliverable. Watch for the weight release, third-party benchmark results, and how quickly Western clouds add Qwen 4 to their catalogs.