Explore Qwen4-Max for coding, professional work, long-running agents, and scientific exploration, with model specifications, API access, explicit caching, and CoT Summary.
Qwen4-Max is the first model in the Qwen4 family, combining reasoning, native vision, and long-running task execution. It supports coding, professional work, design, and scientific exploration, using tools and feedback to turn complex goals into reviewable deliverables.
Model ID:
Coding and engineering — Combines visual understanding, code generation, and professional tools. The launch blog demonstrates building an Unreal Engine game, turning production-line layouts into 3D models, and improving machining plans through simulation feedback.
Professional work — Brings together evidence retrieval, analysis, and deliverable creation. Examples include financial research with linked Excel workbooks and editable presentations, and legal due diligence with document classification, cross-source checks, and traceable citations.
Long-running agents and live delivery — Plans and iterates over extended tasks. With the Qwen Live Delivery plugin, an agent can deliver a usable first version and update it as work progresses, allowing users to review results and redirect the task.
Scientific exploration — Connects research, specialized tools, and experimental feedback. The blog describes a cooling-fluid molecule designed and synthesized through an integrated experimental platform, and drug candidates optimized through computational evaluation for further laboratory validation.
See the launch blog for the examples and full evaluation results.
Streaming, parallel tool calls, prefix completion, and implicit caching are also supported.
Available in Singapore (International), China North 2 (Beijing, China), and Frankfurt, Hong Kong, Tokyo, and US (Virginia) (Global). Beijing and Global share the same prices; Singapore has separate prices.
Fine-tuning and model deployment are not supported. See Pricing and Rate limits. Pay-as-you-go calls in Singapore offer 1 million free tokens, separate from Token Plan credits; see Free quota for eligibility and validity.
Obtain an API Key, set the
Other supported APIs: Responses, Anthropic Messages, and DashScope. Qwen4-Max uses the DashScope multimodal endpoint even for text-only input.
Thinking is enabled by default, with
When
Reuse shared input prefixes such as system prompts, tool definitions, and reference documents to reduce repeated computation, latency, and input costs. Set cache breakpoints manually or let automatic caching place them. A matching prefix can be reused deterministically while its cache remains valid. TTL options are
Return a readable summary of the model's reasoning together with the encrypted full reasoning, instead of returning the full reasoning in plain text. The summary is for display; the encrypted content restores reasoning history in later turns. Enable the feature on every turn and return historical encrypted content unchanged. A summary cannot replace the encrypted reasoning history.
This feature supports Chat Completions, Responses, Anthropic Messages, and DashScope. Summaries or encrypted content can be absent. See CoT Summary for API-specific settings, streaming fields, and multi-turn examples.
Use Qwen4-Max through Token Plan Personal, Token Plan Team, or the pay-as-you-go API. Token Plan usage consumes Credits; use the corresponding API Key and endpoint. The separate
qwen4-max
Key highlights
Coding and engineering — Combines visual understanding, code generation, and professional tools. The launch blog demonstrates building an Unreal Engine game, turning production-line layouts into 3D models, and improving machining plans through simulation feedback.
Professional work — Brings together evidence retrieval, analysis, and deliverable creation. Examples include financial research with linked Excel workbooks and editable presentations, and legal due diligence with document classification, cross-source checks, and traceable citations.
Long-running agents and live delivery — Plans and iterates over extended tasks. With the Qwen Live Delivery plugin, an agent can deliver a usable first version and update it as work progresses, allowing users to review results and redirect the task.
Scientific exploration — Connects research, specialized tools, and experimental feedback. The blog describes a cooling-fluid molecule designed and synthesized through an integrated experimental platform, and drug candidates optimized through computational evaluation for further laboratory validation.
See the launch blog for the examples and full evaluation results.
Capabilities and specifications
| Capability | Support | Capability | Support |
|---|---|---|---|
| Text generation | ✓ | Thinking | ✓ |
| Function calling | ✓ | Structured output | ✓ |
| Image understanding | ✓ | Video understanding | ✓ |
| Explicit cache (Qwen4) | ✓ | CoT Summary | ✓ |
| Web search | ✓ | PDF understanding | ✓ |
| Code interpreter | ✓ | Web scraping | ✓ |
| Text-to-image search | ✓ | Image-to-image search | ✓ |
| MCP | ✓ | Batch inference | Not supported |
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| Context length | 1,000,000 tokens | Maximum output (non-thinking) | 131,072 tokens |
| Maximum input (non-thinking) | 991,808 tokens | Maximum input (thinking) | 983,616 tokens |
| Maximum thinking + final answer | 262,144 tokens | Default thinking budget | 131,072 tokens |
Quick start
Obtain an API Key, set the DASHSCOPE_API_KEY environment variable, and install the OpenAI SDK for your language. The examples below use the pay-as-you-go Chat Completions API. For Token Plan credentials and endpoints, see Token Plan.
Usage notes
Thinking and tool calls
Thinking is enabled by default, with reasoning_effort=high and thinking_budget=131072. Use low, high, or xhigh to select the reasoning effort. In Chat Completions, set enable_thinking=false to turn thinking off. max_completion_tokens limits thinking and the final answer together: up to 262,144 tokens in thinking mode or 131,072 in non-thinking mode.
preserve_thinking is enabled by default. Preserve the assistant messages and their reasoning in multi-turn tool calls. See Thinking and Function calling. The Responses API also supports MCP tool calling.
Default parameters and reasoning effort by API
Default parameters and reasoning effort by API
Parameter | Default |
|---|---|
temperature | 1.0 in thinking mode; 0.7 in non-thinking mode. |
top_p | 0.95 in thinking mode; 0.8 in non-thinking mode. |
top_k | 20 |
repetition_penalty | 1.0 |
presence_penalty | 0.0 (Thinking); 1.5 (Non-thinking) |
logprobs | false |
enable_thinking | true |
preserve_thinking | true |
reasoning_effort | high |
thinking_budget | 131072 |
max_completion_tokens | 262144 (Thinking); 131072 (Non-thinking) |
reasoning_effort and thinking_budget are omitted, the defaults are reasoning_effort=high and thinking_budget=131072.| Protocol | Parameter | Supported values | Default | Mapping |
|---|---|---|---|---|
| OpenAI Chat Completions / DashScope | reasoning_effort | low, high, xhigh | high | minimal → low; medium → high; max → xhigh |
| OpenAI Responses | reasoning.effort | low, high, xhigh | high | minimal → low; medium → high; max → xhigh |
| Anthropic Messages | output_config.effort | low, high, xhigh | high | medium → high; max → xhigh |
Explicit cache (Qwen4)
Reuse shared input prefixes such as system prompts, tool definitions, and reference documents to reduce repeated computation, latency, and input costs. Set cache breakpoints manually or let automatic caching place them. A matching prefix can be reused deterministically while its cache remains valid. TTL options are 5m and 1h, with 5m as the default; implicit cache does not guarantee hits.
Supported APIs are Chat Completions, Responses, and Anthropic Messages. DashScope does not support the new Qwen4 explicit cache. Cache writes and reads use the model's published rates. See Explicit cache (Qwen4) for configuration, renewal, and billing.
CoT Summary
Return a readable summary of the model's reasoning together with the encrypted full reasoning, instead of returning the full reasoning in plain text. The summary is for display; the encrypted content restores reasoning history in later turns. Enable the feature on every turn and return historical encrypted content unchanged. A summary cannot replace the encrypted reasoning history.
This feature supports Chat Completions, Responses, Anthropic Messages, and DashScope. Summaries or encrypted content can be absent. See CoT Summary for API-specific settings, streaming fields, and multi-turn examples.
Use in coding tools
Use Qwen4-Max through Token Plan Personal, Token Plan Team, or the pay-as-you-go API. Token Plan usage consumes Credits; use the corresponding API Key and endpoint. The separate qwen4-max-preview model has its own preview benefits.