Short answer: Qwen4-Max is Alibaba's new flagship AI model and the first model in the Qwen4 family. Alibaba released it on November 16, 2026, after a preview that opened on November 11. It has 2.4 trillion total parameters, of which 71 billion are active for each token, plus native vision and a 1 million token context window (Alibaba Cloud). You can call it today as qwen4-max on QwenCloud, using the OpenAI SDK you probably already have. On QwenCloud's Singapore region it costs $1.50 per million input tokens and $4.50 per million output tokens.
If you saw the launch thread and wondered whether Qwen4-Max is worth wiring into your stack, this guide is for you. It covers what the model is, what the numbers mean, how to call it, what it costs, how it compares with Google's Gemini 4 models, and what "open weights soon" actually promises.
This guide is part of our AI tools for developers map, which breaks down the 2026 ecosystem by category.
What is Qwen4-Max?
Qwen is the family of AI models built by Alibaba. Qwen4-Max is the largest and most capable one so far, and the first released from the new Qwen4 generation. QwenCloud's docs describe it as a model that combines reasoning, native vision and long running task execution, aimed at coding, professional work, design and scientific research (QwenCloud).
"Native vision" means it reads images and video directly, without a separate vision model in front. You can hand it a screenshot, a chart or a recorded demo along with your text.
What do 2.4 trillion and 71 billion parameters mean?
Parameters are the learned numbers inside a model. More of them usually means more knowledge and skill, but also more compute per answer.
Qwen4-Max uses a design called Mixture of Experts. Instead of running every parameter on every word, the model sends each token to a small set of specialist parts. So although Qwen4-Max has 2.4 trillion parameters in total, only about 71 billion do work on any given token. Alibaba says that's 25% fewer active parameters than its predecessor, Qwen3.8-Max, at the same total size, which is how it aims to be stronger and cheaper to run at once.
Specs at a glance
From QwenCloud's model page, checked November 16, 2026:
| Spec | Value |
|---|---|
| Model ID | qwen4-max |
| Context window | 1,000,000 tokens |
| Max input | 991,808 tokens without thinking, 983,616 with thinking |
| Max output | 131,072 tokens without thinking, 262,144 for thinking plus answer |
| Inputs | Text, images, video and PDFs |
| Built in tools | Function calling, web search, web scraping, code interpreter, MCP |
| Not supported | Batch inference, fine tuning, custom deployment |
Thinking is on by default. You can set the reasoning effort to low, high (the default) or xhigh, or turn thinking off in Chat Completions with enable_thinking=false.
How to call Qwen4-Max from your code
The easy part: you don't need a new SDK. QwenCloud accepts OpenAI style requests, so the OpenAI client works once you point it at QwenCloud's endpoint. This example comes straight from QwenCloud's quick start:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.getenv("DASHSCOPE_API_KEY"),
base_url="https://ap-southeast-1.maas.qwencloudapi.com/compatible-mode/v1",
)
response = client.chat.completions.create(
model="qwen4-max",
messages=[{"role": "user", "content": "Create a project plan with staged deliveries."}],
)
print(response.choices[0].message.content)Besides Chat Completions, QwenCloud supports the OpenAI Responses API, the Anthropic Messages API and its own DashScope API. That means you can also use it from coding tools: QwenCloud's docs include setup notes for Claude Code, Codex and OpenCode.
Two things to know before you build on it. First, for multi turn tool calls, keep the assistant messages and their reasoning in the history, since QwenCloud enables preserve_thinking by default. Second, if you resend the same long system prompt or documents, explicit caching can cut repeated input costs. It works with Chat Completions, Responses and Anthropic Messages, with a 5 minute or 1 hour lifetime.
How much does Qwen4-Max cost?
QwenCloud bills per million tokens, in US dollars, with the same price for thinking and non thinking modes and no tiers by input length (QwenCloud pricing, checked November 16, 2026):
| Billing item | Singapore (International) | Beijing and Global regions |
|---|---|---|
| Input | $1.50 | $1.413 |
| Output, including thinking | $4.50 | $4.24 |
| Cache write, 5 minutes | $1.875 | $1.767 |
| Cache write, 1 hour | $3.00 | $2.827 |
| Cache read | $0.075 | $0.071 |
The Global regions are Frankfurt, Hong Kong, Tokyo and US (Virginia). Thinking tokens are billed as output, so a higher reasoning effort means a bigger bill for the same question.
Is there a free option? New pay as you go users in the Singapore region get 1 million free tokens to try the model. Other regions have no free quota. You can also use Qwen4-Max through Alibaba's Token Plan subscription, which runs on credits, or in Alibaba's QwenWork and QwenStudio apps.
What does Alibaba claim about performance?
Alibaba's launch post makes several performance claims. Treat all of these as Alibaba's own numbers, not independent tests:
- Coding: Qwen4-Max ranked 6th on Code Arena, a real world coding leaderboard.
- Efficiency: on Work360, an Alibaba benchmark covering tasks across 721 specialized fields, Alibaba says Qwen4-Max beats Qwen3.8-Max overall while using 35% fewer tokens in its high reasoning mode and cutting the average cost per task by about 60%.
- Agents and science: Alibaba describes a 48 hour continuous agent run that helped discover an experimentally validated coolant molecule for chip cooling.
Benchmarks tell you where to start testing, not what to pick. Run it on a handful of your own tasks before switching a production workload.
Qwen4-Max vs Gemini 4: a quick price check
Qwen4-Max arrived four days after Google's Gemini 4 Hydrogen, and both have a 1 million token window. Here's how list prices compare, checked November 16, 2026:
| Qwen4-Max (Singapore) | Gemini 4 Hydrogen | Gemini 4 Argon | |
|---|---|---|---|
| Input per million tokens | $1.50 | $1.50 | $2, then $4 from December 1, 2026 |
| Output per million tokens | $4.50 | $7.50 | $10, then $20 from December 1, 2026 |
| Free option | 1M free tokens for new Singapore users | Free API tier with rate limits | No free API tier |
| Context window | 1M tokens | About 1M tokens | About 1M tokens |
Sources: QwenCloud pricing and Gemini API pricing. On paper, Qwen4-Max costs the same as Hydrogen for input and less for output. But price per token isn't price per task. A model that thinks longer or needs more retries can cost more overall, so compare on your real prompts. For more on Google's side, read our guides to Gemini 4 Hydrogen and Gemini 4 Argon.
Will Qwen4 be open weight?
Not Qwen4-Max itself, at least for now. It's available only through QwenCloud and Alibaba's apps. What Alibaba said is that "open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide." No dates or sizes were given.
That matters if you want to run a model on your own hardware or fine tune it. Today, Qwen4-Max can't be downloaded or fine tuned, and QwenCloud lists fine tuning as unsupported for it.
Preview vs final model
During the preview, Alibaba offered a separate model, qwen4-max-preview, through its Token Plan only. QwenCloud says the preview runs until November 25, 2026 (UTC+8), and that model will be retired after that. The docs now recommend the production qwen4-max, so if you tested the preview, switch your model ID.
Where Qwen4-Max fits among other model families
Qwen joins a crowded field of strong model families. If you're comparing options, our explainers on Kimi AI and Mistral AI cover two other families developers often test side by side.
The skill behind picking a model
A new flagship model shows up every few weeks. The names change, but the skill stays the same: knowing which model fits a job, testing it on real work and checking what it produces. If you want to get better at building with AI models and agents, at your own pace, the form below is the place to ask how.
Sources, checked November 16, 2026: Alibaba Unveils Qwen4-Max (Alibaba Cloud, November 16, 2026), Latest model: Qwen4-Max (QwenCloud docs), Pricing (QwenCloud docs), qwen4-max-preview (QwenCloud docs), Gemini API pricing (Google).
