Skip to content

Together AI adds Alibaba's Qwen4 Max with 1M-token context

Short answer

Together AI has made Alibaba's Qwen4 Max available on its serverless inference service with a 1M-token context window. Through November 22, 2026 it costs $1.05 per million input tokens and $3.15 per million output tokens. Requests must stream, and thinking is on by default and billed as output.

What this means for operators

If your team handles long documents, such as contract archives, ticket histories or a whole codebase, a 1M-token window lets you send the material in one request without building a retrieval step first. Price a typical job before committing: the vendor's own test, a 652,519-token input, cost 69 cents at the sale rate and 98 cents at list. Check three things in a pilot: the 54-second wait before the first token on a large input, thinking tokens billed as output, and the streaming-only requirement in any existing integration code. Repeat requests on the same material are cheaper because of automatic caching.

Together AI has put Alibaba's flagship Qwen model, Qwen4 Max, on its serverless inference service under the model string Qwen/Qwen4-Max. Alibaba released the model on November 16, 2026. Together says the launch price is 30% below list through November 22, 2026.

What the model is

Qwen4 Max is a sparse mixture-of-experts model with 2.4 trillion parameters, 71 billion of them active per token. It accepts text and image input and returns text. Tool calling works through the standard tools parameter, and a single turn can return more than one call. Alibaba positions it for long agent work, including coding across large repositories. On its own Work360 benchmark, Alibaba reports better overall results than Qwen3.8 Max, 35% fewer tokens at high reasoning effort and roughly 60% lower cost per task. Those are Alibaba's figures, and Together did not verify them.

Pricing

Sale price (to Nov 22, 2026)List price
Input, per 1M tokens$1.05$1.50
Cached input, per 1M tokens$0.0525$0.075
Output, per 1M tokens$3.15$4.50

Prompt caching is automatic but best effort, so not every repeated request is served from cache.

Together's own test

Together sent the model a bug report and all 344 files of the Tornado Python web framework, 652,519 tokens, in one request with no tools and no file selection. The first token came after 54 seconds and the full answer after 72. The patch passed the two regression tests the maintainers wrote for the bug, and Tornado's full suite of 1,673 tests had no failures with the patch applied. The model also pointed to related error-handling code in a file the bug report never mentioned. The request cost about 69 cents at the sale price.

Together chose this as a single example, using a fix the maintainers committed after the model's release. It shows what one call can do, and it is not a benchmark. A repeat of a roughly 630,000-token request came back almost entirely from cache. The first token arrived in 8 seconds and the input cost about 3 cents instead of 66.

Constraints to know before calling it

  • Only streaming requests are accepted. A request without streaming returns a 400 error.
  • Thinking is on by default. The reasoning streams first, so there is a pause before the answer, and thinking tokens count as output tokens. It can be switched off per request.
  • The model is serverless only. Fine-tuning and dedicated endpoints are not available. Alibaba says smaller open-weight Qwen4 models will follow.

The script Together published sends every text file that git tracks, so check a repository for secrets before sending it to any hosted model.

Source: Together AI · In the Atlas: Together AI →

Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.