How do you use Qwen3.8-Max?

Qwen3.8-Max is Alibaba's new flagship model. After a preview in July, the hosted version went live on Alibaba's cloud on August 3, and on August 12 the Qwen team published its weights as Qwen3.8-2.4T-A95B. The name describes the size: 2.4 trillion parameters in total, of which about 95 billion are active for each token. That is a mixture of experts: a router sends each piece of text to a small group of specialist sub-networks, so the model costs far less to run than its total size suggests.

Alibaba presents it as a model for coding, professional work and long, multi-step agent tasks. Its own benchmark table puts it close to Claude Opus 4.8 and in places near Claude Fable 5 and GPT-5.6 Sol, though these are the maker's figures, run in its own test setups.

Three ways to use it

1. Qwen Studio, in the browser. Alibaba's chat site at chat.qwen.ai lets you pick Qwen3.8-Max from the model list. This is the quickest way to judge it on your own questions. The hosted Max version has features the open weights lack: it accepts images, offers a non-thinking mode for quick replies, and keeps up to a million tokens of context by default.

2. The API, through Qwen Cloud. Alibaba's official API service follows the OpenAI format, so the standard OpenAI Python library works. Create a key in Qwen Cloud, copy the base address and the exact Qwen3.8-Max model name shown in your console, and set them as environment variables:

import os
from openai import OpenAI

client = OpenAI(base_url=os.environ["QWEN_BASE_URL"], api_key=os.environ["QWEN_API_KEY"])

reply = client.chat.completions.create(
    model=os.environ["QWEN_MODEL"],  # the Qwen3.8-Max name listed in your console
    messages=[{"role": "user", "content": "Explique en trois phrases ce qu'est un taux d'intérêt composé."}],
    reasoning_effort="medium",
    extra_body={"enable_thinking": True, "preserve_thinking": True},
)
print(reply.choices[0].message.content)

reasoning_effort sets how long the model thinks: xhigh is the default and the most thorough, medium balances speed and accuracy, and low is the fastest and cheapest. preserve_thinking keeps the model's earlier reasoning available in a long conversation, which helps on multi-step tasks.

3. The open weights. Anyone can download Qwen3.8-2.4T-A95B from Hugging Face and serve it with engines such as vLLM or SGLang, but only on server-class hardware. The open model is text-only and always thinks before it answers. Check its licence before building on it: providers earning more than US$50 million within 12 months need a separate commercial licence from Alibaba.

A worked bilingual test

Benchmarks rarely tell you how a model handles your language and your work. A short test does. Pick five real tasks you do every week and write each in French and in English. For example:

Rédige un courriel de deux paragraphes à un client de Gatineau pour reporter une livraison de mardi à jeudi. Ton poli, vouvoiement, pas d'excuses excessives. Termine par une seule question.

Send the French and the English version to each model you want to compare, in a fresh chat each time. Then score every answer from 0 to 2 on four points:

An illustrative result sheet might read: French 7 out of 8, English 8 out of 8 for one model, and 6 and 8 for another. The gap between the two languages for the same model is often more revealing than the totals.

How does it compare with Western models in French?

No one can say yet from published numbers. Alibaba's model card reports coding, agent and research benchmarks, and two of its in-house tests are bilingual in English and Chinese, but it gives no French-specific scores. So any claim that Qwen3.8-Max is better or worse in French than Claude, GPT or Gemini models is, for now, an impression rather than a measurement.

If differences show up, they are most likely at the edges rather than in basic grammar: regional vocabulary, formal register, idioms, and staying in French when the source document is in English. Those are the points the test table above measures, which is why your own five tasks are worth more than a leaderboard here.

Two practical points round out the comparison. First, the hosted version runs on Alibaba's cloud, so apply the same care you would with any provider and keep confidential client data out until you have read its terms. Second, if French quality turns out to matter most and Qwen3.8-Max scores lower than your current model on your test, it can still earn a place for coding or research tasks, with another model kept for client-facing writing.