Alibaba opens Qwen4-Max-Preview to developers for two weeks only
Alibaba Cloud published documentation for Qwen4-Max-Preview, its first Qwen4-generation model, available until November 25.
· Source: Associated Press · Last verified: 12 Nov 2026

Summary
- Alibaba Cloud published documentation for Qwen4-Max-Preview, its first Qwen4-generation model, available until November 25.
- The model carries a 1-million-token context window and runs only on paid Token Plan subscriptions.
- It works through OpenAI and Anthropic interfaces, letting developers test coding and agent workloads before general release.
The latest
Alibaba Cloud has opened its first Qwen4-generation model to developers through a two-week preview window that closes at 24:00 Beijing time on November 25, 2026. Documentation on Model Studio, last updated November 11, describes qwen4-max-preview as the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. The model will not be modified during the preview, and is withdrawn afterwards.
Details
- The specs: Alibaba Cloud lists a context length of 1 million tokens and a maximum output of 393,216 tokens shared between reasoning traces and the final answer. The documentation does not state a separate input ceiling. The model accepts text, images and video for screenshot analysis, document understanding and video analysis.
- The target workloads: According to the documentation, the model is aimed at coding, office productivity, front-end interface design and complex agentic tasks — the categories where Chinese labs have pushed hardest against US rivals. Alibaba has not published benchmarks, parameter counts, or a model card for the preview.
- The reasoning controls: Thinking is enabled by default and tuned through a reasoning_effort parameter with low, medium and xhigh settings, xhigh being the default. Alibaba Cloud warns that reasoning_effort cannot be combined with thinking_budget in the same call.
- The token trap: In the Chat Completions interface, max_tokens equals max_completion_tokens here, so reasoning consumes part of the same budget and a limit sized only for the answer may truncate output. Earlier Qwen models keep their previous behavior. preserve_thinking is on by default, with prior reasoning placed in reasoning_content.
- The interfaces: The model supports function calling, web search and built-in tools, with structured outputs via JSON Object and JSON Schema. Compatibility covers OpenAI Chat Completions, OpenAI Responses and Anthropic Messages — meaning existing codebases built against either stack can point at Alibaba's endpoint with minimal rewriting.
- The access path: Alibaba Cloud says the model runs on the Token Plan service only, with no pay-as-you-go option. Developers need a dedicated API key from Token Plan Personal or Token Plan Team, routed through a Singapore-region base endpoint at token-plan.ap-southeast-1.maas.aliyuncs.com.
- The pricing gap: The documentation states no per-token price and no credit consumption rate, deferring to separate Token Plan guides. Neurohive reported that during the preview the model draws credits at no less than 20% of the standard rate, falling to 8% between 22:00 and 08:00 Beijing time.
- The subscription tiers: Neurohive put Token Plan Lite at $6 monthly, discounted to $3 for new subscribers until November 30 Beijing time, with Essential at $10, Standard at $18 and Pro at $43. Data Studios reported team plans ranging from $20 to $180 per seat.
- The announcement: Neurohive reported that the Qwen team announced the model at 02:01 UTC on November 11, 2026 with a short post on X and Weibo reading Qwen4 is coming, while the model was already live via QwenWork, Qoder, the Qwen app and QwenStudio. Data Studios listed an alternate identifier, qwen4.0-max-preview.
Between the lines
Shipping behind OpenAI and Anthropic interfaces lowers the switching cost to near zero for developers already on those stacks, which is the clearest read on who the preview is built for. The absence of benchmarks, a parameter count or a model card, paired with a fixed two-week window and subscription-only access, suggests a capacity-constrained trial rather than a launch.
What's next
The preview closes at 24:00 Beijing time on November 25, 2026, after which the model is withdrawn. The new-subscriber discount runs until November 30. Watch for whether Alibaba publishes benchmarks or a model card at general release.