On November 21, 2026, Together AI announced that Alibaba's flagship model Qwen4 Max is live on its Serverless Inference platform under the model string Qwen/Qwen4-Max, with a 1-million-token context window; a launch discount of 30% runs through November 22. Together also put it through a whole-repository test: it fed the model a bug report plus all 344 files of the Tornado repository in a single request, and 72 seconds later the patch it returned passed the maintainers' tests.
The test gave the model no tools, preselected no files, and ran no retrieval step; the 344 files totaled 652,519 tokens, and the model had to locate the problem itself inside the million-token context. On price, the discount brings input to $1.05 per million tokens, cached input to $0.0525, and output to $3.15; against QwenCloud's official prices of $1.50 for input and $4.50 for output, that is about 30% cheaper.
Qwen4 Max is the flagship of the Qwen4 series that Alibaba released on November 16, a 2.4-trillion-parameter MoE architecture with native vision and a million-token context; its official channel is QwenCloud, and Together is a third-party inference platform, so this listing gives developers one more place to call it. Two details matter before calling it: requests must stream, and thinking is on by default, with thinking tokens billed as output tokens, so the real bill depends on how long the model thinks; the model is closed-weight and API-only, with no local deployment.
For developers who need whole-repository bug fixes or long-document review, this is a practical option that swallows an entire codebase in one request, at a price 30% below the official channel. But the 30% discount ends on November 22, so anyone who wants to try it should move fast; and the preview model string qwen4-max-preview retires on November 25, so calls still on the preview need to switch to the production qwen4-max.