Models

Qwen3.8-Max and Open Qwen Models: Which One Should You Choose?

Qwen is Alibaba’s model family and one of the most price-competitive alternatives to Western flagship models. In 2026, its lineup includes Qwen3.8-Max, with 2.4 trillion parameters and a one-million-token context window at $2/$6, as well as open-weight models you can run yourself. This guide covers the lineup, licensing caveats, and how Qwen handles Russian.

Current Qwen models

Qwen3.8-Max launched in August 2026 after its presentation at WAIC in Shanghai. Its unusual mixture-of-experts (MoE) design has 2.4 trillion parameters, but only about 95 billion are active for a given request. That helps keep inference prices manageable despite the model’s size.

Alibaba also releases open weights. Qwen3.8-27B is a dense model that can run on a consumer GPU with 24 GB of memory. Alibaba says it competes with the much larger Qwen3.7-Plus, making it a strong option for local use and experimentation.

ModelWhat it isPrice / license
Qwen3.8-MaxFlagship: 2.4T parameters (MoE), 1M context, text and video$2 input / $6 output; closed API
Qwen3.8-27BOpen dense model that runs on one graphics cardApache 2.0; free
Qwen3.8-Flash-NextPreview of the architecture planned for Qwen4Qwen Community License
Qwen3.5 / 3.6 / 3.7Earlier generations still in useApache 2.0 for open versions

How capable is it?

The flagship is positioned for long-running agent tasks such as multi-day autonomous coding, reproducing research, and coordinating several agents. Russian-language performance has traditionally been a Qwen strength: the model was trained on substantial Russian-language data and often performs well against Western models in the same price range.

Qwen3.8-Max natively accepts video and images. One caveat: the open flagship variant (2.4T-A95B) is text-only; multimodal input is available only through the closed API version.

Safety and refusals

Qwen is trained under Chinese regulations. Political topics are filtered more strictly than by many Western models, and this also affects the open weights. Refusals are less common for everyday and work tasks, where the model tends to answer a broad range of technical questions.

Open weights give you more control: you can run the model locally, remove the service layer, and avoid sending data to a third party. They cannot erase what the model learned, though; baseline refusals on political topics remain in the weights.

What does it cost?

At $2/$6 per million tokens, the flagship is in Grok’s price range and 2.5 to 15 times cheaper than GPT Sol or Claude Fable. Cached reads cost $0.25. For large-scale Russian text processing, Qwen can offer strong value for money.

Local use has no per-token fee, but requires suitable hardware. A 27B model in 4-bit quantization uses about 17 GB of memory, so plan for a 24 GB GPU or a Mac with enough unified memory. Without that hardware, an API version or an aggregator’s lower-cost model may be simpler.

Key features and limitations

  • Open models have a native 262K-token context window, extendable to 1M with YaRN; the flagship supports 1M out of the box.
  • Newer models enable Thinking by default. You can disable it in the request and control reasoning depth with `reasoning_effort`.
  • Licenses vary. Most open models use Apache 2.0, while flagship weights have a separate license with revenue-related terms. Review the license before commercial use.
  • The family also includes models for code (Qwen3-Coder), images (Qwen-Image), speech, and even driving (Qwen-Drive).
  • Releases are frequent: versions 3.5 through 3.8 arrived within six months. Check the current model page instead of relying on older reviews.

Who should consider Qwen?

If you work in Russian and have a limited budget, Qwen is a good first model to try: it combines strong Russian-language support with low prices. Developers may value the open weights for integrating a model into their products without per-token API costs. Qwen is not the only major Chinese model family: Moonshot AI takes a different approach with Kimi, so it is useful to compare the two.

Organizations with strict requirements around Western compliance and audits may prefer GPT or Claude, since Alibaba operates under a different jurisdiction and data-handling framework. Browse the model catalog for Qwen versions currently available through API providers, or try a model in the chat.

Frequently asked questions

Is Qwen free?

Open models are free to download and run. The Qwen3.8-Max API flagship costs $2/$6 per million tokens. The chat at chat.qwen.ai offers free access with limits.

Can I run Qwen on my computer?

Yes. Open versions range from 0.6B to 27B parameters. The 27B model needs about 24 GB of video memory in 4-bit form.

Why is Qwen so inexpensive?

Its MoE architecture activates only a small share of its parameters per request, and Alibaba is pricing aggressively to gain market share.

Is Qwen censored?

Political topics are filtered more heavily than by many Western models. Refusals are less common on technical and everyday topics.

How does Qwen compare with Kimi?

Both are Chinese model families with open weights and low prices. Kimi (Moonshot) is usually more expensive and stronger at agent tasks; Qwen has a broader lineup and stronger Russian-language support. See the Kimi review.

Sources

How we check our guides