Alibaba released Qwen4-Max on November 16, 2026. It is the first model in the Qwen4 family and the successor to Qwen3.8-Max for coding, visual reasoning and tool-using workloads. The formal release opens pay-as-you-go access alongside Token Plan. ToolWorthy reviewed Alibaba's documentation and launch data for this page. Alibaba launch announcement
What Changed
- Lower active parameter count. Alibaba reports 71 billion active parameters per token, down from 95 billion in Qwen3.8-Max. Total parameters remain at 2.4 trillion. Both Max models accept text, images and video, use a 1M-token context window and return text. Alibaba launch announcement · Qwen3.8-Max reference
- Qwen4 explicit caching. For workloads that reuse long prompts or tool definitions, QwenCloud documents automatic or manually placed cache breakpoints with a five-minute default TTL and a one-hour option. The Qwen4 explicit cache is available with Chat Completions, Responses and Anthropic Messages. DashScope does not support this cache feature. QwenCloud Qwen4-Max reference
- Formal API access and a different reasoning budget. Use
qwen4-maxwith Chat Completions, Responses, Anthropic Messages or DashScope. Thinking is enabled by default athigheffort. The formal release allows up to 262,144 tokens for thinking and the final answer together, or 131,072 output tokens with thinking off. QwenCloud Qwen4-Max reference
Compared With Qwen3.8-Max
| Decision | Qwen3.8-Max | Qwen4-Max |
|---|---|---|
| Input and output | Text, image and video in. Text out. | Text, image and video in. Text out. |
| Singapore input and output per million tokens | $2.00 input, $6.00 output | $1.50 input, $4.50 output |
| Singapore implicit cache hits per million input tokens | $0.25 | $0.075 |
| Batch inference | QwenCloud lists batch support for the qwen3.8-max alias in Beijing |
Unsupported on QwenCloud |
QwenCloud publishes these rates in USD. If you rely on batch jobs, check the region and exact model ID before changing the route. Qwen3.8-Max reference · Qwen4-Max API reference · QwenCloud pricing
Alibaba-reported Work360 results
On Alibaba's internal Work360 benchmark, Alibaba reports that Qwen4-Max performed better overall than Qwen3.8-Max, used 35% fewer tokens at High reasoning effort, and lowered average cost per task by about 60%. These are Alibaba's own results for that workload. Cost on your requests will depend on token usage, cache hits and region. Alibaba's Work360 results
Migrating From Qwen4-Max-Preview
The qwen4-max-preview model was released on November 11 and is scheduled to retire when the preview ends at midnight at the end of November 25, 2026 (UTC+8). Switch the model ID to qwen4-max. If you also switch from Token Plan to pay-as-you-go, use the API key and base URL for your chosen billing route and region. Qwen4-Max Preview reference · QwenCloud Qwen4-Max reference
- Recheck long outputs. Preview allows up to 393,216 tokens for thinking and the final answer combined. The formal release caps the combined budget at 262,144. Check completion and truncation on your longest prompts.
- Set reasoning effort intentionally. Preview defaults to
xhigh. The formal release defaults tohigh. On Chat Completions, setreasoning_effortexplicitly if your workload depends on that setting. Responses and Anthropic Messages document their own parameter names. Preview settings · Formal release settings
Access and Pricing
Alibaba lists the model in QwenStudio and QwenWork. Developers can use QwenCloud Token Plan Personal or Team for credits-based usage, or the metered API. Alibaba announced plans to release smaller open-weight Qwen4 models. The announcement described Qwen4-Max access through hosted products and APIs. Alibaba launch announcement · QwenCloud access guide
| Pay-as-you-go region | Input | Output, including thinking | Cache read |
|---|---|---|---|
| Singapore (International) | $1.50 | $4.50 | $0.075 |
| Beijing or Global deployment | $1.413 | $4.24 | $0.071 |
Prices are in USD per million tokens. Cache writes have separate rates for five-minute and one-hour TTLs. These are per-token API rates. Token Plan charges against Credits under its own terms. QwenCloud pricing · QwenCloud access guide
Who Should Upgrade or Wait
- Move off Preview if you depend on continued access. Recheck output caps and quality at the reasoning effort you will use.
- Pilot Qwen4-Max for interactive coding and visual analysis jobs already routed to Qwen3.8-Max. Compare completed-task cost, tool-call behavior and response time on your own prompts.
- Keep an existing route for Beijing batch jobs until you have a verified replacement. For native audio input, keep a supported Omni route such as Qwen3.8-Omni-Flash.

