Qwen AI Models Guide: Qwen3.8 Max, Pricing & Polish Test

👁 126 reads

Qwen is Alibaba’s family of language models: from small ones you can run on your own GPU to Qwen3.8 Max with 2.4 trillion parameters, which costs $2 per million input tokens via API. You can chat with it for free at chat.qwen.ai and in the Qwen Android app. We tested it in Polish - it writes correctly and handles error correction better than other Chinese models, though it can spend the entire token budget thinking and return nothing.

This guide covers language models. The image model has a separate guide: Qwen-Image-2.1.

What is Qwen

Qwen (in Chinese: Tongyi Qianwen) consists of models developed since 2023 by the Qwen team at Alibaba Cloud. The family is broad: text models in many sizes, multimodal models (Omni), code models, speech models, a live translator, and the Qwen-Image generator. Most smaller models have open weights on Hugging Face, which is why Qwen is one of the most downloaded open model families and the base for thousands of derivative models.

For users, there are four entry points:

• Qwen Chat / Qwen Studio (chat.qwen.ai) - free browser-based chat with model selection,
• Qwen app - on Android “Qwen - Alibaba AI Assistant” from Qwen Team (over 5 million downloads, in-app purchases); we couldn’t find an official app in the App Store,
• Qwen Cloud API / Alibaba Cloud Model Studio - pay per token, international region in Singapore,
• Local models - weights from Hugging Face run via Ollama, llama.cpp, or vLLM.

How to get started step by step

1. Open chat.qwen.ai and create an account. Select the Qwen3.8 Max model from the list and write in Polish - it responds in the language of the question.
2. On Android, install the “Qwen - Alibaba AI Assistant” app from Google Play. Check the publisher (Qwen Team) - there are many similar-named wrappers in the stores.
3. For API access, the easiest route is through OpenRouter: model qwen/qwen3.8-max-0902, with Alibaba as the only provider. Alternatively, create an account in Alibaba Cloud Model Studio (international version, card payment), where you have native enable_thinking and thinking_budget parameters.
4. Set a thinking limit. When writing in Polish, use low level and set a generous response limit (16,000 tokens or more) - otherwise you might pay for an empty response.
5. For code, install Qwen Code - an open agent in the terminal (Apache 2.0). In PowerShell: irm https://qwen-code-assets.oss-cn-hangzhou.aliyuncs.com/installation/install-qwen-standalone.ps1 | iex, then qwen in your project directory and /auth to provide your provider and key. It supports Qwen, OpenAI, Anthropic, Gemini APIs, and local models from Ollama.
6. Locally: Qwen3.8 27B in Q4_K_M quantization takes 17 GB of GPU memory and generates about 50 tokens per second on RTX 4090. Instructions: Ollama and local AI.

Qwen models in October 2026

Qwen3.8 Max in the API is a version with extras: image input, no-thinking mode, default 1 million token context, and built-in tools. The open weights Qwen3.8-2.4T-A95B is the same base model but text-only. The license is similar to MIT with two conditions: a product with over 100 million monthly users or $20 million monthly revenue must display the model name, and a company with a “model as a service” offering or code/office assistant with over $50 million annual revenue needs a separate license. There are no conditions for internal use.

OpenRouter also has Qwen3.8 Max Prime (since September 23): the same model served faster, at $4 / $12, twice as expensive. We wrote about the launch and benchmarks in Qwen3.8-Max: everything we know and where it wins and where it loses.

Benchmarks

In Alibaba’s table, Qwen3.8 Max scores 92.6% on GPQA Diamond, 86.6% on Terminal-Bench 2.1, 67.7% on SWE-bench Pro, and 82.8% on IFBench (instruction following), where it beats GPT-5.6 Sol (72.7%) and Claude Fable 5 (63.5%). These are manufacturer numbers. In the independent Artificial Analysis index, the 0902 snapshot scores 45 points, same as GLM-5.3 and one point more than Kimi K3. The same site shows a weakness: 39 tokens per second and $5.41 per index task, the most in the top tier, because the model generates very large amounts of thinking tokens.

How much does Qwen cost

Prices from OpenRouter’s model list, October 3, 2026. Qwen3.8 Max costs the same at Alibaba (Model Studio, Singapore region, cache read $0.25), and Alibaba doesn’t split pricing by context length tiers. On OpenRouter, Qwen3.8 27B also has a free variant with 262k token context. The free chat at chat.qwen.ai and downloading open weights are free. The Android app has in-app purchases, but Alibaba doesn’t publish pricing for chat plans.

Qwen in Polish

Qwen3.8 Max writes correctly in Polish and follows instructions very precisely. In our September test of 13 tasks in our Lab, it scored 76.4 out of 100 points (68.2 with equal token limits) and was the only Chinese model to meet all 26 rules. Details in Qwen3.8 Max in Polish. In our October test:

• it corrected all 14 errors in proofreading and as a bonus removed the pleonasm “wróciłem z powrotem“ (came back back),
• it answered 20 questions about Poland flawlessly, including the current president (Karol Nawrocki) and the two capitals of Lubusz Voivodeship,
• it translated idioms naturally (“wziąć byka za rogi“ - take the bull by the horns, “złapmy się w poniedziałek“ - let’s catch up on Monday), but left “onboarding” untranslated,
• in an email it suggested “Thursday, October 3” - that was a Saturday.

The weakness is thinking. Even at low level, Qwen thought on every task: 17.8 thousand reasoning tokens across eight short queries, over 8 times more than Kimi K3. Worst with exact numbers and text length - in September the Prime version counted every word in its thoughts and didn’t manage to respond.

Our test: Qwen3.8 Max vs Kimi, GLM and DeepSeek

On October 3, 2026, we gave Qwen and four other Chinese models the same six tasks in Polish via OpenRouter, with low reasoning, 8000 token limit, and Alibaba as provider. Full description in Kimi K3, GLM-5.3, Qwen3.8 Max and DeepSeek in Polish: test.

The empty response concerned three questions with dates. Qwen thought for 171 seconds, used all 8000 tokens and wrote nothing, and we paid $0.048. On a retry with a 16,000 limit, it needed only 2,212 thinking tokens and 57 seconds, and all three answers were correct. If we counted the retry, Qwen would have 57 out of 60 points and would have won the test. However, we count the first attempt, as with all models.

Qwen vs the competition

Guides to the others: Kimi, GLM, DeepSeek. A shorter September introduction is in Qwen AI: what it is.

Who is Qwen for

• For proofreading and tasks with hard rules - Qwen follows instructions most precisely among Chinese models.
• For running locally - Qwen3.8 27B on Apache 2.0 is one of the two best local models in Polish in our RTX 4090 test (72.2 points, just behind Gemma 4 12B).
• For multimodal tasks - Omni Flash accepts audio and video at $0.15 per million input tokens.
• Not where response time matters. 448 seconds for eight queries was the slowest in our test.
• Not for sensitive data in chat - data goes to Alibaba’s servers. The local version sends nothing. In the US, there’s ongoing discussion about banning Chinese models, including Qwen.

Frequently asked questions

Qwen - what is it?

Qwen is Alibaba’s family of AI models: chat.qwen.ai chat, mobile app, API, and open models to download. The newest flagship is Qwen3.8 Max from August 2026, with the 0902 snapshot from September.

Is Qwen free?

The chat.qwen.ai chat is free, and you can download open models from Hugging Face without fees. The API is paid: Qwen3.8 Max costs $2 per million input tokens and $6 per million output tokens.

Does Qwen work in Polish?

Yes. In our test, Qwen3.8 Max corrected 14 out of 14 errors and answered 20 questions about Poland flawlessly. It writes correctly, though sometimes leaves anglicisms like “onboarding” untranslated.

How much does Qwen3.8 Max cost?

$2.00 per million input tokens and $6.00 per million output tokens, with no context length tiers. The Prime version costs twice as much.

Is Qwen open source?

Most models are. Qwen3.8 27B has an Apache 2.0 license, and Qwen3.8-2.4T-A95B, the base Max-class model, has a proprietary license with restrictions only for large companies. The Qwen3.8 Max API version has extras not present in the open weights.

Qwen or DeepSeek?

In our test, DeepSeek V4.1 Flash scored 93.3 points and Qwen3.8 Max 82.5, mainly due to Qwen’s empty response. Qwen follows rules better and has small models for home use, DeepSeek is cheaper and doesn’t freeze on thinking.

Is Qwen safe?

Models run locally don’t send data. Chat and API process data on Alibaba’s servers. For company documents, choose the local version or an EU-based provider.

How to run Qwen locally?

Easiest via Ollama or llama.cpp. Qwen3.8 27B in Q4_K_M quantization needs about 17 GB of GPU memory - fits on RTX 4090 and 3090.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and verification date

• Parameters
• licenses
• and manufacturer benchmarks: model cards Qwen3.8-2.4T-A95B and Qwen3.8-27B on Hugging Face and LICENSE file. Launch dates: Qwen Cloud model changelog and model card on OpenRouter. Pricing: OpenRouter model list and Alibaba Cloud Model Studio pricing. Index and speed: Artificial Analysis (read September 27
• 2026). App: Google Play and App Store search. Qwen Code: QwenLM/qwen-code repository on GitHub. Test results are our measurements via OpenRouter (October 3
• 2026) and locally on RTX 4090 (September 26
• 2026). Verified October 3
• 2026

Kinetyra