How to Use the Qwen3.8 Max API

Qwen3.8 Max is available on EmpirioLabs. It is the flagship tier of Alibaba's Qwen3.8 series, a trillion-scale mixture-of-experts model built for coding, long-horizon agent work, and professional tasks, with a 1M token context window, deep thinking, and text, image, and video input. It runs on an OpenAI-compatible API, so you can switch to it by changing one field.

What Qwen3.8 Max is good at

Max is the capability tier of the Qwen3.8 family, so it suits work where quality matters more than raw speed: coding across a large repository, long-running agent loops, document and screenshot understanding, and analysis that has to hold a lot of context at once. It reads images and video, calls tools with standard function calling, returns strictly schema-conformant JSON on request, and can search the web, run code, read URLs, and search for images through built-in tools.

How to call it

Send a standard chat completion:

curl https://api.empiriolabs.ai/v1/chat/completions \
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-8-max",
    "messages": [
      {"role": "user", "content": "Refactor this module and explain the tradeoffs."}
    ]
  }'

The same model id also works on /v1/responses, on the Anthropic-shaped /v1/messages, and on the Google-compatible /v1beta/models/qwen3-8-max:generateContent route, so most existing SDK setups work without a rewrite.

Thinking is on by default

Qwen3.8 Max reasons before it answers unless you turn that off. Set enable_thinking to false for short, latency-sensitive calls, or raise thinking_budget up to 262144 tokens when you want it to work a problem harder. Reasoning tokens are billed as output tokens, so a large budget on a simple prompt is the main way to overspend on this model.

One practical detail worth knowing before you port an agent over: while thinking is enabled, the model accepts tool_choice only as auto. Forcing a specific function with tool_choice: "required" or an explicit tool object is rejected. If your agent forces a first tool call, either let the choice stay automatic or disable thinking for that step.

Built-in tools default to on

Five built-in tools ship with the model: web search, web extractor, code interpreter, text-to-image search, and image-to-image search. They are all enabled by default, which is convenient for research-style prompts and wasteful for a plain completion, because the tool definitions ride along in every prompt. Turn off the ones you do not need with the tool_* parameters:

{
  "model": "qwen3-8-max",
  "messages": [{"role": "user", "content": "What changed in the spec this week?"}],
  "tool_web_search": true,
  "tool_web_extractor": true,
  "tool_code_interpreter": false,
  "tool_web_search_image": false,
  "tool_image_search": false
}

Two dependencies are enforced upstream: web extractor only runs alongside web search, and both web extractor and code interpreter require thinking to be enabled. Web search and the two image search tools are billed per invoked call, and a single request can invoke a tool more than once, so the response reports exactly what ran in usage.tool_usage.

Structured output

Qwen3.8 Max enforces a strict JSON Schema, not just JSON mode, so you can ask for an exact response shape and parse it without a repair step:

{
  "model": "qwen3-8-max",
  "messages": [{"role": "user", "content": "Extract the invoice fields."}],
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "invoice",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "vendor": {"type": "string"},
          "total": {"type": "number"}
        },
        "required": ["vendor", "total"],
        "additionalProperties": false
      }
    }
  }
}

Images and video

Attach images or video the same way you would with any OpenAI-compatible vision model, using image_url or video_url content parts. You can send several images in one message and refer to them positionally. For video, video_fps controls how densely frames are sampled, and vl_high_resolution_images trades detail against prompt size on image input.

Pricing

Qwen3.8 Max is pay as you go, priced per input and output token, plus a small per-call fee on the search tools when they actually run. Thinking tokens count as output tokens. Current rates are on the Qwen3.8 Max model page and the pricing page, which always reflect the live rate card.

Try it

Open it in the EmpirioLabs Playground to test prompts, attachments, and the tool toggles without writing code, or read the API reference for the full parameter list.