Qwen 4 Max by Alibaba
Alibaba’s Qwen 4 Max is a flagship multimodal reasoning model with 1M-token context, image and video understanding, and tool calling for agent workflows.
Overview
Qwen 4 Max is Alibaba’s flagship Qwen4 model for deep reasoning and long-horizon tasks. It uses a sparse Mixture-of-Experts design with 2.4T total parameters and 71B active parameters per token. It can read text, images, video, and PDFs, then produce a text answer you can ship as a deliverable. It also supports tool calling, so it can plan steps, call functions, and continue work across turns.
What you can build
- Code generation and refactors over large repos with up to 1,000,000 tokens of context
- Agent workflows that call your tools for search, retrieval, and task execution
- Screenshot-to-code for frontend prototypes, UI audits, and bug triage
- PDF Q&A for research papers, contracts, and internal documentation
- Form and document extraction into structured JSON for downstream automation
- Video understanding for summaries, event timelines, and timestamped action lists
Inputs
- A required instruction or question as plain text.
- Optional images to ground the request. On Wiro, you can attach up to 10 images, up to 10 MB each.
- Optional multi-turn chat history as a list of messages with roles. Use this instead of a single prompt when you need a conversation.
- Optional system instructions that set behavior and constraints for the whole run.
- Optional reasoning depth selection. You can turn reasoning off or ask for deeper reasoning.
- Optional web search enablement. When enabled, the model may search the web during the run.
- Optional function tools you define for tool calling. Provide names, descriptions, and JSON input shapes.
- Optional tool routing guidance. You can let the model decide, forbid tools, or require tool use.
- Optional chat identifiers (user and session IDs) to keep history consistent across runs.
- Optional continuation inputs for tool workflows. You can pass a prior run token, plus tool results, to resume.
- Optional output length cap to limit total generated tokens, including reasoning.
Outputs
The model returns a chat-style result in JSON.
- The primary output is an assistant message in text.
- If you enabled tools, the model may return one or more tool-call requests with JSON arguments.
- If you requested structured output, the assistant message can be valid JSON that matches your schema.
- When reasoning is enabled, providers may return extra reasoning metadata such as summaries for display.
Limitations
- Qwen 4 Max is a closed-weights model. You access it through hosted APIs, not downloadable weights.
- Context is capped at 1,000,000 tokens. Output is capped at 262,144 tokens when reasoning is on.
- Batch inference is not supported for this model on QwenCloud.
- Fine-tuning and dedicated model deployment are not supported on QwenCloud for this model.
- Visual inputs have strict limits. Very large image sets, large Base64 payloads, or oversized requests can fail.
- Videos have size and duration limits. Frame sampling can miss brief events in fast motion.
- PDFs have hard limits. Very large PDFs should be passed by URL, not Base64, because Base64 expands size.
- Structured JSON can become less strict with visual inputs on some providers.
- Low-quality inputs reduce accuracy. Scanned PDFs, blurry screenshots, extreme aspect ratios, and messy tables can degrade results.
Safety & compliance
Alibaba states it applies safety controls across training and post-training. It reports added code security checks during training to reduce vulnerable patterns. It also reports filtering cybersecurity data by intent, keeping defensive material while removing offensive attack content. Alibaba also reports extra safeguards for chemical, biological, radiological, and nuclear risk areas.
Don’t use this model to generate instructions for wrongdoing, malware, weaponization, or illicit surveillance. Remove sensitive personal data from prompts when you can. For regulated domains, add human review and logging before you act on outputs.
API quick start
Run qwen 4 max with a single API call.
{
"prompt": "How to roast coffee beans at home?",
"inputImage": "https://your-cdn.com/input.png",
"user_id": "...",
"session_id": "..."
}curl -X POST "https://api.wiro.ai/v1/Run/alibaba/qwen-4-max" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "How to roast coffee beans at home?",
"inputImage": "https://your-cdn.com/input.png",
"user_id": "...",
"session_id": "..."
}
JSONOther models you might like
Explore similar llm & chat models.



