Qwen 3.8 Max is built on a mixture-of-experts (MoE) architecture: 2.4 trillion parameters and a context window of up to 1 million tokens. Next to it came Qwen3.8-27B, a dense model with 27 billion parameters. Its weights are open under the Apache 2.0 license, and one decent graphics card is enough to run it. The chat for the whole lineup is called Qwen Studio and opens at chat.qwen.ai. Below: how the models differ, what the API costs and when it makes sense to host the open version yourself.
The Qwen 3.8 lineup: Max, 27B and Flash
Alibaba Cloud Model Studio describes Qwen3.8-Max as a flagship MoE model with 2.4 trillion parameters, tuned for code and office work such as law, finance and design. According to the official documentation the context window reaches 1,000,000 tokens. The maximum input is just under 992 thousand, and output is capped at 131,072 tokens. The current snapshot is called qwen3.8-max-0902 and was updated on September 2, 2026, while the first version of the lineup shipped in early August.
You cannot call Max a closed model. The QwenLM/Qwen3.8 README says that on August 12, 2026 Qwen3.8-2.4T-A95B was published on Hugging Face and ModelScope. It has 2.4 trillion parameters in total and 95 billion active, a native context of 262,144 tokens (extendable to 1,010,000) and text-only input. According to the model card, Qwen3.8-Max is the official version built on these weights: it accepts images, can run without reasoning, keeps a 1M context by default and ships with built-in tools.
Its license is not Apache 2.0. The metadata says license: other, and the file is named "Qwen3.8-Max License". It is an MIT-style text with two conditions. If a commercial product has more than 100 million monthly active users or revenue above $20 million a month, it must show the model name in the interface. If you sell access to the model as a service (Model as a Service), or build an AI assistant for code and office work and your revenue over 12 months exceeds $50 million, you need a separate license from Qwen.
The open Qwen3.8-27B appeared on Hugging Face in August 2026 under Apache 2.0. Its card calls it the most capable open model in the Qwen family to date. It understands images and video, lets you tune how deeply it reasons and plans agent steps better. The context is 262,144 tokens natively and up to 1 million with extension. The model is dense, not MoE, so it does not need a cluster: a single GPU will do.
Qwen3.8-Flash is meant for bulk work where Max quality is overkill: classification, summarization, short agent steps. It is much cheaper.
Seen a bare "Qwen Max" in a query? In 2026 that means the current Qwen3.8-Max. Older versions count as outdated for new projects.
Qwen Studio: the official chat and site
The official Qwen site is qwen.ai. The web chat is called Qwen Studio and runs at chat.qwen.ai. It offers both open and proprietary models from the lineup, there is nothing to install locally, and signing up takes a couple of minutes.
For API development the Qwen3.8 README puts QwenCloud (qwencloud.com) first. Keys are issued and billing is handled there, and OpenAI and Anthropic compatibility is named as the main way to connect. The second route goes through Alibaba Cloud Model Studio, which has its own regional price list, and some of Alibaba's examples are still written for it.
Qwen 3.8 API pricing
All prices are in dollars. The data comes from the QwenCloud model pages as of October 29, 2026:
| Model | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| Qwen3.8-Max | $2 | $6 |
| Qwen3.8-Flash | $0.15 | $0.47 |
The Qwen3.8-Max page in Alibaba Cloud Model Studio also lists dollar prices, but they differ by region. Singapore has the same $2 and $6, Beijing charges $1.65 for input and $4.951 for output. That is about 17% cheaper, not 60-70%. The Chinese region, though, requires an account and billing inside mainland China. Caching and batch processing are priced differently on each platform, so check the QwenCloud and Model Studio pages before you work out a budget.
Alibaba also runs a Coding Plan, a subscription for agentic coding that replaces pay-per-token billing. The Pro plan costs 50 dollars a month and includes up to 90 thousand requests. The Lite plan has been closed to new subscribers since March 20, 2026, and for people who signed up earlier, renewals and upgrades were switched off on April 13, 2026.
Open weights or API: which to choose
If your data must not leave your own infrastructure, or you want to decide which model version runs, take Qwen3.8-27B and deploy it yourself. The 2.4-trillion 2.4T-A95B will not run on a single graphics card. For a quick start without hardware, the QwenCloud or Model Studio API is simpler, or the Qwen Studio chat.
If payment from Russia is the sticking point, Qwen is in better shape than most Western services. Nobody has announced a direct IP block, and the barrier is usually verifying an Alibaba Cloud account and getting a foreign card accepted. Claude is the opposite: a premium model with access from Russia limited by Anthropic's geography. codexkey closes that gap: Claude Opus 5.5 and Claude Sonnet 5, paid with a Russian card or through SBP (the Russian Faster Payments System), with no Anthropic account and no VPN. In the catalog on October 29, 2026, Sonnet 5 costs $0.52 per 1M input tokens and $2.6 for output, Opus 5.5 $0.86 and $4.32. Qwen3.8 models are there too: qwen3.8-max, qwen3.8-flash, qwen3.8-omni-flash and the free qwen3.8-flash-free. Prices are on the models page. Where an open model wins and where premium Claude does is covered in Claude vs Qwen.
If you want a coding agent built on Qwen, we covered installing and using Qwen Code separately, including how it compares with Claude Code.
FAQ
Where is the official Qwen site?
The main site is qwen.ai. The chat for everyday work is called Qwen Studio and opens at chat.qwen.ai. API keys are issued on QwenCloud (qwencloud.com) and in Alibaba Cloud Model Studio.
How does Qwen3.8-27B differ from Qwen3.8-Max?
Max is an MoE model with 2.4 trillion parameters, the official version built on the open Qwen3.8-2.4T-A95B weights, with image input and built-in tools. It is easiest to use through the API or the chat. 27B is an open dense model with 27 billion parameters under Apache 2.0: you can download it and run it on your own graphics card, but on hard tasks it falls behind the flagship.
How much does Qwen 3.8 Max cost in the API?
On QwenCloud and in the Singapore region of Model Studio: $2 per 1 million input tokens and $6 per million output tokens. The Beijing region of Model Studio is slightly cheaper ($1.65 and $4.951), but it needs a Chinese account.
Is Qwen 3.8 free?
The open models in the lineup, Qwen3.8-27B included, can be downloaded for free. If you run them yourself, you pay only for hardware. The Qwen Studio chat is free but limited. The paid API is billed per token.
What is Qwen Studio?
It is the name of Alibaba's web chat for the Qwen lineup, available at chat.qwen.ai. Do not confuse it with QwenCloud and Model Studio: those are developer platforms where you issue API keys and set up billing.
Does Qwen work from Russia?
There is no official statement about a block, and no direct IP ban has been published. But verifying an Alibaba Cloud account and paying with a foreign card often get in the way for users in Russia.