Hopp til hovedinnhold

Alibaba releases Qwen4-Max in preview – first Qwen 4 model targets coding and agents

Qwen4-Max is now available in preview through Alibaba's Token Plan subscription and in four of its own apps. Per-token pricing, benchmarks and licence are not disclosed.

Av Joachim HøgbyKilde: Qwen

Alibaba has released the first model in its new Qwen 4 family. The flagship Qwen4-Max is now available in preview through the Token Plan subscription, and selected users can try it in four of Alibaba's own apps. The company is aiming the model at coding, knowledge work and AI agents that carry out multi-step tasks.

The Qwen team announced the launch on its official X account on Wednesday 11 November at 03:01 Oslo time (02:01 UTC). The post opens with "Qwen4 is coming", and the rest of the family has not launched yet.

What Alibaba says about the model

According to the Qwen team, Qwen4-Max has deeper knowledge, a richer understanding of the world and stronger coding and agentic capabilities. The model is built for real-world tasks, the team writes, naming complex coding, knowledge work, front-end builds and professional software workflows.

These are Alibaba's own descriptions. The announcement contains no benchmark results, no information on the model's size or how much text it can handle at once, and no comparison with models from OpenAI, Anthropic or Google. Instead, the team asks users to "put it to the test on a real problem" and show what they build.

Alibaba has said before that Qwen 4 was in training, as hogby.ai reported in September (in Norwegian). Now the first model is out of the lab, but only as a preview.

How to get access

Qwen4-Max sits in Token Plan, a monthly subscription from Alibaba's cloud service Qwen Cloud. The plan gives single access to several models for text, vision, speech and image generation, including models from other vendors. The page shows models from Zhipu (GLM) and DeepSeek among others.

Token Plan comes in four individual tiers: Lite, Essential, Standard and Pro. Essential gives 2.25 times the credits of Lite, Standard four times and Pro ten times. The number of agents that can run at the same time rises from one to two on Lite to six to eight on Pro. During a promotion, new customers can get the plan from 3 dollars a month, while Lite and Essential normally start at 6 dollars, according to the page. The final price is set at checkout.

The plan works with tools that support the OpenAI and Anthropic protocols. Alibaba lists Qwen Code, Cline, Claude Code, Cursor, OpenCode, Codex, Kilo CLI and OpenClaw. The Qwen team also offers early access in QwenWork, the coding tool Qoder, the Qwen app and QwenStudio.

What is not disclosed

The announcement does not give a per-token price in the regular API, how much of the plan's credit a Qwen4-Max request uses, or when the model becomes generally available. Nor does it say whether Qwen4-Max or other models in the family will get open weights, as several earlier Qwen models have, or which licence applies.

The Token Plan page states that the service is provided by Intelligent Cloud Computing (Singapore) Private Limited. Where data is processed, and whether processing happens inside the EEA, is not stated in the announcement.

What business leaders should know

Qwen is Alibaba's most important model family, and in our assessment a new Max model makes the race for coding agents even tighter. For European businesses that means more options, but also more questions.

  • Test it, but don't build on it yet. Qwen4-Max is a preview without published benchmarks or a fixed price. Try it on a well-defined problem and compare it with the tools you already pay for.
  • Check where your code ends up. When developers connect Cursor, Claude Code or Codex to a new subscription, code and data go to a new vendor. Settle privacy, the data processing agreement and storage location before anyone uses the model on company code.
  • Factor in the security findings. In October Cohere published findings that an earlier Qwen model tried to send a test file out of a closed sandbox, as hogby.ai reported. That says nothing about Qwen4-Max, but it underlines that agents with tool access should be tested in isolated environments.

Sources and media

Illustration: AI-generated with Grok Bot / hogby.ai

Mer om AI

Oppskriften. Strategien. Hver dag.

AI-nyheter for ledere. Kuratert av en CIO som bygger det selv. Daglig i innboksen.