Key takeaways
- Qwen4-Max, released on November 16, 2026, is available only through Alibaba Cloud and QwenCloud APIs, with no downloadable weights.
- Alibaba says smaller Qwen4 models are coming to open source but has given no sizes, dates or license.
- Qwen3.8-27B is open source under Apache 2.0, while the open Qwen3.8 Max-class and Flash-Next models use custom Qwen licenses with commercial conditions.
No, Qwen4-Max is not open source. Alibaba released it on November 16, 2026 as a hosted model only: you can call it through Alibaba Cloud and QwenCloud, but there are no weights to download, and no Qwen4 model of any size was on Qwen's Hugging Face page as of November 16. Alibaba says smaller Qwen4 models will get open weights, without saying when.
That is a change from the last generation. In August, Alibaba published the weights of the model behind Qwen3.8-Max, the first time it had opened a Max-class model. Anyone who expected Qwen4-Max to follow the same path will have to wait, or settle for a smaller sibling.
Is Qwen4-Max open weight
No. Every official page treats it as a service:
- The Model Studio docs list one model ID,
qwen4-max, reached through the OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope APIs. The same page marks "model deployment" as not supported, so you can't even run a private copy inside Alibaba Cloud. - The QwenCloud model page has prices, rate limits and code samples for the API, and nothing about weights or a license.
- A search of Hugging Face for models from the Qwen organization named "Qwen4" returned nothing on November 16. The newest Qwen language models there are still the Qwen3.8 family from August.
- Artificial Analysis, which tests models independently, lists Qwen4-Max as proprietary, with Alibaba Cloud as its only API provider.
Bloomberg reported the same on launch day: Techmeme's summary of its story says Qwen4-Max "isn't open weight" and that Alibaba plans to release weights for smaller Qwen4 models.
What Alibaba has said about open Qwen4 models
The Qwen4-Max launch post says smaller models in the Qwen4 family "are coming to open source", built with the same training methods as Max. It gives no sizes, no dates and no license. The post itself only open-sources a plugin, Qwen-Live-Delivery-Plugin.
So the honest answer for anyone planning around Qwen4 weights is: some are promised, none are out, and the terms are unknown. That last point matters, because Qwen's recent licenses vary a lot (see below).
The launch post also spends more space than usual on safety. It says Alibaba filtered Qwen4-Max's training data by purpose, keeping material for defensive security work such as finding and fixing vulnerabilities and removing material meant to help attacks. It also says it added dedicated safety data for chemical, biological, radiological and nuclear risks. The South China Morning Post led its launch story with these guardrails. Alibaba hasn't said whether this is part of why Max stays closed. It is a fair guess, though: anyone can fine-tune public weights to add back what was filtered out, while a hosted model stays under the company's control.
Which Qwen models are open source, and under what license
"Open weight" means the trained model files are free to download and run. "Open source" in the strict sense also means a standard license, such as Apache 2.0, with no extra conditions. Qwen's recent models fall on both sides of that line. Here is what the license files on Hugging Face said on November 16, 2026:
| Model | Released | Weights on Hugging Face | License |
|---|---|---|---|
| Qwen4-Max | November 2026 | No, API only | None published |
| Smaller Qwen4 models | Not yet | Promised, no date | Not announced |
| Qwen3.8-2.4T-A95B (base of Qwen3.8-Max) | August 2026 | Yes | Qwen3.8-Max License |
| Qwen3.8-Flash-Next | August 2026 | Yes | Qwen Community License 1.0 |
| Qwen3.8-27B | August 2026 | Yes | Apache 2.0 |
| Qwen3.6-35B-A3B | April 2026 | Yes | Apache 2.0 |
| Qwen3.6-27B | April 2026 | Yes | Apache 2.0 |
The three licenses work differently:
- Apache 2.0 (Qwen3.8-27B, Qwen3.6-27B, Qwen3.6-35B-A3B). Use, change, host and sell it, as long as you keep the copyright notice. This is open source by any definition.
- Qwen3.8-Max License (Qwen3.8-2.4T-A95B). It allows the same uses, with two conditions. A product with more than 100 million monthly active users or $20 million in monthly revenue must show the model name prominently. And a company running a "Model as a Service" business (selling API access) or an "AI Work Assistant" (a coding or office assistant) needs a separate license from Qwen once its revenue passes $50 million over 12 months. Internal use is exempt.
- Qwen Community License 1.0 (Qwen3.8-Flash-Next). Same display rule, but stricter on the second point: the text sets no revenue threshold, so any Model as a Service or AI Work Assistant business needs a separate license from Qwen for commercial use.
Even the open Max-class model isn't the same as the hosted one. The Qwen3.8-2.4T-A95B model card says the official Qwen3.8-Max adds vision input, a non-thinking mode, a 1 million token context by default and built-in tools. For a full breakdown of that generation, see is Qwen 3.8 open source.
How to use Qwen4-Max instead
Through the API. QwenCloud lists it at $1.50 per million input tokens and $4.50 per million output tokens as of November 16, 2026, with cached input at $0.075 per million. It has a 1 million token context window and takes text, image and video input. Because it speaks the Anthropic Messages format, tools built for Claude can point at it by changing the base URL and model name. Our Qwen4-Max launch report covers the price, the regions and Alibaba's benchmarks in detail.
How it compares with the best open models
Closed doesn't mean out of reach for open models. Artificial Analysis gives Qwen4-Max 43 on its Intelligence Index, its overall score from a set of independent tests, which ranks it 38th of 81 models. Artificial Analysis also says Qwen4-Max costs about $3.27 per task on its index.
That gap makes the closed weights matter less than they might. A team that needs a model it can host, inspect or fine-tune already has an open option that scores higher on this index. Qwen4-Max's case rests on price and its long context, and those come with the API.
What to watch next is the first open Qwen4 release: its size, and whether it ships under Apache 2.0 like Qwen3.8-27B or under one of Qwen's custom licenses.
Frequently asked questions
Is Qwen4 on Hugging Face?
Not yet. As of November 16, 2026, the Qwen organization on Hugging Face had no Qwen4 models; its newest language models are the Qwen3.8 family from August 2026.
When will open Qwen4 models be released?
Alibaba has not given a date. The Qwen4-Max launch post says smaller models in the family are coming to open source, without naming sizes or a license.
Is Qwen 3.8 Max open source?
Its base model, Qwen3.8-2.4T-A95B, has open weights on Hugging Face under the Qwen3.8-Max License, which adds conditions for very large products and for API and coding-assistant businesses. The hosted Qwen3.8-Max adds vision, a non-thinking mode and built-in tools that the open weights lack.
Can I run Qwen4-Max locally?
No. There are no weights to download, and Alibaba Cloud Model Studio lists model deployment as not supported, so the only way to use it is the hosted API.
Sources
- Qwen4-Max: Your New Teammate Has ArrivedQwen (Alibaba)
- qwen4-maxAlibaba Cloud Model Studio
- Qwen4-Max model pageQwenCloud
- Qwen organization on Hugging FaceHugging Face (Qwen)
- Qwen3.8-2.4T-A95B model cardHugging Face (Qwen)
- Qwen3.8-2.4T-A95B LICENSEHugging Face (Qwen)
- Qwen3.8-Flash-Next LICENSEHugging Face (Qwen)
- Qwen3.8-27B model cardHugging Face (Qwen)
- Qwen3.6-35B-A3B model cardHugging Face (Qwen)
- Qwen3.6-27B model cardHugging Face (Qwen)
- Qwen4 Max: Intelligence, performance and price analysisArtificial Analysis
- AI models leaderboardArtificial Analysis
- Alibaba unveils new flagship AI model in riposte to US rivalsBloomberg
- Alibaba says new flagship AI model cuts task costs 60% while tightening cyber guardrailsSouth China Morning Post


