- 5 min read
- 8 key claims
- 1 source cited
- For Developers, Businesses
Overview
Alibaba released Qwen4 Max on November 16, 2026, and Together AI made it available through its Serverless Inference API under the model identifier Qwen/Qwen4-Max on November 21. The model is a sparse mixture-of-experts system with 2.4 trillion total parameters and 71 billion active per token. It accepts a context window of up to 1 million tokens, allowing a large codebase or document set to be sent in a single request.
Together AI marked the launch with a 30% discount through November 22, 2026. The promotional rates are $1.05 per million input tokens, $0.0525 per million cached input tokens, and $3.15 per million output tokens. Requests must stream, and thinking is enabled by default, with thinking tokens billed as output tokens [1].
Why it matters
The repository-scale demonstration shows a different approach to code repair than file-selection or retrieval pipelines. The model received the whole project and identified the two relevant files itself, including a precedent the bug report did not mention. Its patch passed the maintainers' regression tests and all 1,673 cases in the full suite [1].
For repeated work over stable material, Together AI's automatic prompt caching changes the cost and latency profile. Resending a request of roughly 630,000 tokens produced an input cost of about 3 cents and a first-token time of 8 seconds, against 66 cents and 54 seconds for the original 652,519-token request [1].
The API constraints matter for application design. Streaming is mandatory, and reasoning tokens are billed as output, so integrations must support streamed responses and budget for output that includes thinking tokens [1].
Opportunities
- Repository maintenance services: The Tornado case indicates that a single API call can handle whole-codebase bug diagnosis and patch generation without a separate retrieval stage [1].
- Long-context assistants: Businesses could apply the 1M-token window to internal documentation, policy archives, and large codebases without manually dividing them into fixed-size chunks [1].
- Repeated-query workflows: Because cached input is priced separately and returned much faster in the demonstrated case, services built around recurring queries against stable context may see lower unit costs than one-off requests [1].
- API-based agent products: The model is reachable through Together's serverless API and Python SDK, so repository-scale repair, migration, and document-analysis workflows can be assembled as hosted services [1].
Risks
Cost predictability is constrained by how thinking tokens are billed. They count as output tokens, and the promotional output rate is $3.15 per million tokens, so workloads that generate long reasoning traces can cost more than the input price implies [1].
Streaming is required, which excludes applications without streaming support. The demonstrated 54-second first-token time on a large uncached prompt also raises questions for interactive use, though typical-prompt latency is not characterized [1].
Pricing after November 22, 2026 is not disclosed, which complicates long-term budgeting [1].
Dependency extends beyond the model: access is through Together AI's serverless infrastructure under the Qwen/Qwen4-Max identifier, and the caching behavior behind the measured cost and latency improvements is tied to that platform [1].
Context
The release targets workloads that exceed ordinary prompt sizes. Together AI's demonstration used the Tornado Python web framework: a bug report plus all 344 repository files, totaling 652,519 tokens, in one request. No tools were provided, and no relevant files were selected or retrieved beforehand [1].
The test was run against the repository state immediately before a fix the maintainers committed on November 18, 2026. The model traced the failure to WebSocketProtocol13.write_message in tornado/websocket.py, then produced a patch that marked the exception as retrieved. It also identified an existing precedent in tornado/iostream.py, where BaseIOStream.write handled the same class of future in the same way [1].
Alibaba positions the model for long-horizon agent work, including coding across large repositories. On its Work360 benchmark, which covers industry tasks across 721 specialized fields, Alibaba reports that Qwen4 Max outperforms Qwen3.8 Max overall, uses 35% fewer tokens at high reasoning effort, and costs about 60% less per task on average. These are vendor-supplied figures rather than independent evaluations [1].
Known, claimed, inferred
The evidence supports a narrow conclusion about practical value: Qwen4 Max can consume a complete repository and produce a fix that satisfies maintainer-written tests. It does not establish that this performance generalizes across repositories, bug classes, or languages [1].
Automatic prompt caching appears likely to favor recurring work over stable context rather than one-off requests, since the repeated Tornado request was served mostly from cache and showed both lower cost and lower latency. This is an inference from one example rather than a measured pattern [1].
Repository-scale code repair also appears closer to a near-term fit than latency-sensitive interactive assistance, based on the single demonstrated workflow and the first-token time recorded for an uncached large prompt [1].
Limitations & uncertainties
Only one release and one official Together AI article are represented, so the record contains no independent confirmation of the release details [1].
There are no third-party benchmark results for Qwen4 Max and no comparison with competing models such as GPT-4 or Claude [1].
The evaluation covers a single repository and a single bug-fix scenario, with no distribution of results across repositories, languages, or task types [1].
No user testimonials or broader real-world usage reports are present [1].
Tool support, retrieval integration, and structured-output behavior are not documented beyond the no-tools, no-retrieval conditions of the one example [1].
It is not established whether the cached-input rate applies identically across all Together AI plans or workloads, and no multi-request concurrency or rate-limit information is given [1].
Non-sale input and output rates, along with any pricing changes after the promotion ends on November 22, 2026, are not provided [1].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementQwen4 Max is available on Together AI's Serverless Inference API under the identifier `Qwen/Qwen4-Max`, with a 1M-token context window.[1]
- Primary-source or vendor statementThe model is a sparse mixture-of-experts system with 2.4 trillion total parameters and 71 billion active per token.[1]
- Primary-source or vendor statementIn Together AI's test, Qwen4 Max received all 344 files of the Tornado repository plus a bug report in one request, then returned a patch that passed the maintainers' tests and all 1,673 cases in the full suite.[1]
- Primary-source or vendor statementThat 652,519-token request cost 69 cents at sale price, with first-token latency of 54 seconds and a complete reply at 72 seconds.[1]
- Primary-source or vendor statementA repeat request of roughly 630,000 tokens was served mostly from cache, cutting input cost to about 3 cents and first-token latency to 8 seconds.[1]
- Primary-source or vendor statementThe promotional rates through November 22, 2026 are $1.05 per million input tokens, $0.0525 per million cached input tokens, and $3.15 per million output tokens.[1]
- Primary-source or vendor statementRequests must stream, and thinking is on by default with thinking tokens counted as output tokens.[1]
- Primary-source or vendor statementAlibaba reports that Qwen4 Max outperforms Qwen3.8 Max on Work360, uses 35% fewer tokens at high reasoning effort, and costs about 60% less per task on average.[1]
Sources
- 1Qwen4 Max is live on Together AI●together.airetrieved Nov 21, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.