Alibaba launches Qwen4-Max with 2.4 trillion parameters and 1 million token context
Alibaba's new flagship Mixture-of-Experts model activates 71 billion parameters per token, offers a 1 million token context window and direct image input, while cutting average task costs by 60 percent.

By TechWhoKnows
3 min read
Alibaba launched Qwen4-Max on 16 November 2026, positioning the model as its most capable release to date and a direct competitor to frontier models from OpenAI, Anthropic and Google.
The model uses a Mixture-of-Experts architecture with 2.4 trillion total parameters. Alibaba says it activates 71 billion parameters per token, 25 percent less than its Qwen3.8-Max predecessor, which enables the claimed 60 percent reduction in average task cost.
Key specifications include a context window of up to 1 million tokens and direct image input, allowing the model to process screenshots, scanned documents, charts and UI states without a separate OCR step. The model is accessible through QwenCloud (API), Qwen Studio, QwenWork and Qoder, Alibaba's coding-focused surface.
The release marks a strategic shift for the Qwen family: Qwen4-Max ships without open weights. Earlier Qwen releases were known for open-weight availability, which let researchers inspect expert routing and measure active parameters directly. With a hosted-only model, the parameter count and performance claims become vendor statements that cannot be independently verified.
Reported API pricing, attributed to Bloomberg rather than an official Alibaba price page, sits at $1.50 per million input tokens and $4.50 per million output tokens. If accurate, the input price matches Google's Gemini 4 Hydrogen tier, suggesting Alibaba is competing on capability and task cost rather than undercutting on sticker price.
Alibaba says it tightened safety controls around secure coding, defensive versus offensive cyber material, and chemical, biological, radiological and nuclear (CBRN) risk. The company has not published an independent audit of these claims. The limited preview that opened on 11 November retires on 25 November, giving developers a short window to migrate.
For enterprise buyers, the combination of long context, native vision and lower per-task pricing targets agent workloads where prompts grow with every step. The closed-weight decision means teams needing data-residency or self-hosted deployment must stay on older open Qwen releases or evaluate alternatives.
Sources
- 1Shattered.io analysisshattered.io
- 2Unite.AI coverageunite.ai