📅 November 16, 2026 • alibabacloud.com
Alibaba Cloud has announced Qwen4-Max, a new flagship model the company’s blog positions as an “AI teammate for work, creativity and discovery.” The model carries 2.4 trillion total parameters, of which roughly 71 billion activate per token under a sparse mixture-of-experts design — per the company, that is about 25% fewer activated parameters than its predecessor, Qwen3.8-Max. Sparse activation is the mechanism behind the efficiency figures: on any single pass most parameters stay idle, so serving cost tracks the active set rather than the full parameter count.
Long context and native vision. **Per the announcement, Qwen4-Max accepts a **1M-token context window and can return up to 393K output tokens in a single response. Input is natively multimodal — text, images and video are handled by the same model rather than bolted-on adapters. The company frames the combination as built for long-document analysis and video-heavy work alongside text.
Cost per task and pricing. **On **Work360, a vendor-internal benchmark spanning 721 specialized fields, Alibaba reports roughly 60% lower average cost per task versus Qwen3.8-Max. The predecessor averaged RMB 26.58 per task on the same suite. The suite is vendor-built, so these are self-reported comparisons. Pricing starts at $0.075 per 1M input tokens — per the company, that tier implicitly corresponds to cache-hit requests. Independent coverage from MarkTechPost, TechNode and Business Standard ran the same day.
What it means in practice. **The headline capability for builders is long context: a full codebase, a contract archive or hours of transcripts can sit in a single window instead of a retrieval layer. Native image and video input removes a preprocessing stage from document and media pipelines — invoices, screenshots and clips go straight into the model. For high-volume automation the economics are quoted per task, not per token: at roughly **60% lower cost per task (vendor-reported) and pricing from $0.075 per 1M input tokens, batch workflows get cheaper without architecture changes.
**Availability. **Qwen4-Max is served through an API compatible with OpenAI and Anthropic formats, so existing client code can be repointed with minimal changes, and default rate limits are published in the announcement. Open-weight Qwen4 variants are promised but not yet released — for now the flagship is API-only.
Sources:
alibabacloud.com/blog/alibaba-unveils-qwen4-max-new-ai-teammate-for-work-creativity-and-discovery_603783
www.marktechpost.com
technode.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

