Vibeleaderboard
← All Intel
Intel / article

Qwen4 Max is live on Together AI

Source
together.ai
Date
Why it matters

Qwen4 Max is callable now with 1M tokens of at stated prices. Requests must stream and thinking tokens bill as output, which affects cost planning for whole-repo coding tasks.

Key takeaways · AI-distilled
  • Together AI sent Qwen4 Max all 344 files of the Tornado repository (652,519 tokens) and a symptom-only bug report, with no tools or . In 72 seconds it returned a patch that passed both regression tests the maintainers wrote for that bug. Together calls it one example.
  • Qwen4 Max traced the bug to WebSocketProtocol13.write_message, which wraps each write in a task whose failure nobody has to collect. Its fix marks that exception as retrieved, the same approach BaseIOStream.write takes in tornado/iostream.py, a file the bug report never named.
  • In Together's test, repeating a roughly 630,000-token request came back almost entirely cached: the first token arrived in 8 seconds instead of 54, and input cost about 3 cents instead of 66. Together says caching is best effort, so not every repeat is cached.
  • On Together AI, Qwen4 Max accepts requests only (a non-streaming call returns a 400 streaming_required error). Thinking is on by default and thinking tokens are billed as output; reasoning={"enabled": False} turns it off.
  • Alibaba describes Qwen4 Max as a sparse model with 2.4T total and 71B active parameters. On its own Work360 benchmark, Alibaba reports it beats Qwen3.8 Max overall while using 35% fewer tokens at high effort and costing about 60% less per task.
Terms in this piece · Glossary
  • streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
  • Context and context window — Context is the information available for a generation; the context window is the capacity available to hold it.
  • Retrieval — Selecting evidence relevant to the current request.
  • mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
Read the source www.together.ai
Recommended reads
Comments

Checking sign-in…

Loading comments…