홈 블로그

Qwen4 Max API 사용 방법

API를 통한 Qwen4 Max

Nov 16, 2026

EmpirioLabs AI

고지: 이 글은 AI의 도움을 받아 작성되었으며 EmpirioLabs AI가 검토했습니다.

Qwen4 Max는 2026년 11월 16일에 출시된 알리바바 Qwen 계열의 차세대 플래그십입니다. 텍스트 생성, 이미지 및 비디오 이해, 심층 추론을 하나의 모델로 결합했으며, 코딩, 사무실 생산성, 프론트엔드 디자인, 복잡한 에이전트 작업을 위해 설계되었습니다. 최대 100만 개의 맥락 토큰을 읽고, 기본적으로 답변 전에 생각하며, 세 가지 추론 노력 수준으로 그 사고를 노출시킵니다.

Qwen4 Max는 OpenAI 호환 API를 통해 EmpirioLabs에서 제공되며, 함수 호출, 엄격한 JSON 스키마 구조화 출력, 스트리밍, 그리고 사고 기능이 켜진 최대 262,144개의 출력 토큰을 지원합니다. 다음에서 시도해 보세요. playground 또는 어떤 OpenAI 호환 클라이언트에서든 호출할 수 있습니다. 전체 사양과 현재 요금은 Qwen4 Max 모델 페이지·.

Watch the launch film

Every program in the film was written by Qwen4 Max in one shot through the EmpirioLabs API: a self-playing piano, a starship HUD and a weather app. The score was generated by ACE-Step on EmpirioLabs. Watch it on YouTube.

가격

Usage based. One flat input rate and one flat output rate apply at every prompt size, and thinking tokens bill as output tokens. Cached input. Repeated prompt prefixes are cached automatically, and cached input tokens bill at a steeply discounted rate, so long system prompts and multi-turn conversations cost less on every turn after the first. Built-in tools. Web search, text-to-image search, and image-to-image search add a small fee for each call that actually runs; the web extractor and code interpreter run at no extra cost. Current rates are on the model page and the pricing page.

Quickstart

Point any OpenAI SDK at the EmpirioLabs base URL and pass qwen4-max as the model:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.empiriolabs.ai/v1",
    api_key="YOUR_EMPIRIOLABS_API_KEY",
)

resp = client.chat.completions.create(
    model="qwen4-max",
    messages=[
        {"role": "user", "content": [
            {"type": "text", "text": "Rebuild this landing page section as responsive HTML and CSS."},
            {"type": "image_url", "image_url": {"url": "https://example.com/mockup.png"}},
        ]},
    ],
    reasoning_effort="xhigh",
)

print(resp.choices[0].message.content)

The same model id works on /v1/responses, the Anthropic-shaped /v1/messages, and the Google-compatible /v1beta/models/qwen4-max:generateContent route.

Thinking and reasoning effort

Thinking is on by default and streams back as reasoning_content alongside the answer. reasoning_effort sets how much the model reasons before it answers: low, high (the default), or xhigh. Higher effort spends more reasoning tokens. For a hard token ceiling on reasoning instead, set thinking_budget (up to 262,144 tokens), which runs at the default effort. Set enable_thinking to false for short, latency-sensitive calls.

Built-in tools

Five built-in tools sit behind tool_* parameters and are all off by default: tool_web_search, tool_web_extractor, tool_code_interpreter, tool_web_search_image, and tool_image_search. The web extractor requires web search, and both the extractor and the code interpreter run only while thinking is on. When a tool runs, the response carries a usage.tool_usage map with the number of calls, so you can audit every billed call.

Structured output and function calling

For an exact response shape, pass response_format with {"type": "json_schema", ...} and "strict": true: the model returns exactly the schema's keys. Plain JSON mode with {"type": "json_object"} is also supported. Function calling works with thinking on, including tool_choice: "required" and a named function, and the model can call several functions in one turn.

Things worth knowing before your first call

  • max_tokens covers reasoning and the answer together. With thinking on, a small cap leaves little room for the answer once the model has reasoned. Leave room above the reasoning you expect, or lower reasoning_effort for short replies.
  • Choose either a reasoning effort or a thinking budget. low and xhigh cannot be combined with thinking_budget in the same request; a budget on its own runs at the default effort.
  • JSON mode needs the word "json" in your messages. A {"type": "json_object"} request is rejected unless the word appears somewhere in the conversation. JSON Schema requests have no such requirement.
  • Prompt caching is automatic. Explicit cache controls such as cache_control are not supported and return an error; send the same prefix and the cache applies on its own.

Qwen4 Max is available now in the playground and through the EmpirioLabs API.

더 나은 엔드포인트를 사용할 준비가 되셨나요?

저희 모델을 탐색하시거나, 비즈니스 문의, 맞춤형 배포 또는 기타 어떤 사항이든 문의해 주세요.