Qwen 3.8 Max

Qwen 3.8 Max (qwen/qwen3.8-max) is Alibaba’s new flagship: controllable-depth reasoning, image understanding, a 1M-token context. Benchmark-topping, strong at code and math. The catalog publishes both cache rates — read and write — so the savings of long sessions with a repeating prefix can be calculated up front.

Parameters and capabilities

Parameter	Value
model	qwen/qwen3.8-max (or the short qwen3.8-max)
Context	1M tokens
Reasoning	always on; depth is controllable
reasoning_effort	minimal · low · medium · high · xhigh; off is not supported
Image input	yes — image_url with a data: URL
File input	no — a file part returns 400
Cache read	28.1 ₽ per 1M tokens
Cache write	280.6 ₽ per 1M tokens

Reasoning depth is set by reasoning_effort: five steps from minimal to xhigh. Thinking cannot be turned off completely — the catalog declares no off for this model; the lowest step minimal keeps the token spend under control.

Examples

A code task with a chosen depth

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.mixen.ai/v1",
    api_key="mxn-...",
)

resp = client.chat.completions.create(
    model="qwen/qwen3.8-max",
    reasoning_effort="high",
    messages=[{
        "role": "user",
        "content": "Refactor this function: remove duplication, preserve behavior. <code>",
    }],
)
print(resp.choices[0].message.content)

cURL

curl https://api.mixen.ai/v1/chat/completions \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.8-max",
    "reasoning_effort": "high",
    "messages": [{
      "role": "user",
      "content": "Refactor this function: remove duplication, preserve behavior. <code>"
    }]
  }'

Reading an image

{
  "model": "qwen/qwen3.8-max",
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "What is wrong with this chart? Walk through the anomalies."},
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KG..."}}
    ]
  }]
}

Context caching in /v1/chat/completions kicks in automatically on a repeating prefix: a stable system prompt in a long session is first written for 280.6 ₽ and then read for 28.1 ₽ per 1M tokens.

Pricing

Per 1M tokens: input 224.5 ₽ ($2.5), output 673.4 ₽ ($7.5), cache read 28.1 ₽ ($0.312), cache write 280.6 ₽ ($3.125). Current prices — catalog.

Strengths and limitations

• The flagship of the Qwen 3.8 line — code, math, benchmark-topping tasks.
• Reasoning depth is controllable: minimal for cheap runs, xhigh for the hard ones.
• Understands images — the only non-text input; files (PDF/DOCX) are not accepted.
• A 1M-token context — long documents and whole repositories fit entirely.
• Both cache rates are published — read and write: long-session savings are calculable up front.
• Reasoning cannot be disabled: off is not supported; thinking tokens count toward the spend and eat into max_tokens — set a generous ceiling for long answers.

All models — in the catalog. The general workflow for text models — the chat guide.