Yes for vision and web front ends, less so for agents. Qwen3.8-Max ranks second on the Arena vision board and seventh for web development, which is excellent. But independent agentic tests place it well down the table, and it is one of the slowest and costliest models per task in its group. Its open-weight siblings are often the better buy.
First principles
A model can lead one test and trail another. Each board measures one kind of work. The Arena vision board covers image understanding, the Arena web development board covers building web pages, and this review reads the Vals Index as a test of multi-step agentic work. Strength on one says little about the others, and Qwen3.8-Max shows exactly that split.
Time and cost per task both scale with tokens. The time a task takes is the tokens the model writes divided by its output speed. The cost is those tokens multiplied by the price. A model that is slow and also writes many tokens loses twice: users wait longer and the bill grows. Qwen3.8-Max’s rate card of $2 input and $6 output looks moderate; its measured cost per task does not.
The region is part of the product. A hosted model’s features depend on the region that serves it, so the same model name can mean different capabilities in Beijing, Singapore or Frankfurt.
Free weights are not free of conditions. Downloading open weights costs nothing, but the licence sets the terms of use, and some add conditions for large products. The Qwen3.8 siblings below use three different licences, from Apache 2.0 to a custom licence with such conditions, so read the one for the model you ship.
What it is
Qwen3.8-Max is the flagship of Alibaba’s Qwen3.8 generation. It is available only through Alibaba’s application programming interfaces (APIs). It reads text, images and video and returns text, with a context window of 1,000,000 tokens and up to 131,072 output tokens. It supports a thinking mode.
Sources disagree on its release date. Qwen Cloud’s changelog lists the launch as 3 August 2026, while Alibaba Cloud’s model page shows 28 September 2026, and a snapshot dated 2 September also exists. We leave the date out.
The open-weight siblings are the reason many developers follow Qwen:
- Qwen3.8-2.4T-A95B has 2.4T total and 95B active parameters. It is text only, and thinking cannot be switched off. It uses a custom licence, the Qwen3.8-Max License.
- Qwen3.8-27B is a dense vision-language model under Apache 2.0, with a native context of 262,144 tokens.
- Qwen3.8-Flash-Next has 125B total and 6B active parameters and takes text, images and video. It uses a Qwen community licence whose terms we did not read.
Alibaba also shipped real-time omni models in September, including qwen3.8-omni-flash-realtime on 21 September 2026.
Where it is good
Vision. On the Arena vision board (updated 28 September 2026) Qwen3.8-Max ranks second with 1301, behind only Claude Fable 5 and ahead of every GPT and Gemini entry in the top ten.
Web development. On the Arena web development board (updated 29 September 2026) it ranks seventh at 1671, the highest-placed model from a vendor outside the United States (US).
Breadth. Qwen offers open weights from a 27B dense model up to a 2.4T mixture-of-experts model. Artificial Analysis gives Qwen3.8-27B a score of 34 on its open-weights page, a strong result for its size.
Privacy terms. Alibaba Cloud’s privacy page says Model Studio will not use customer data for model training.
Where it is bad
Agentic work. The Vals Index (updated 29 September 2026) gives Qwen3.8-Max 48.27%, 22nd of 27. That is below DeepSeek-V4.1-Flash and GPT-6 Luna, both far cheaper.
Speed and cost per task. These Artificial Analysis figures are for the snapshot it lists as Qwen3.8 Max (0902), which may differ from the newest release. It measured about 39 output tokens per second, among the slowest near its score, and a cost of $5.41 per index task. That is close to Claude Opus 5.5 at max effort ($5.98), for an index score of 45 against 58. Vals shows a similar pattern at $4.96 per test.
Regional gaps. The model page says fine-tuning is not supported, batch inference runs only in the Beijing region, and web search is disabled in the Frankfurt, US, Japan and Hong Kong regions.
Licence traps in the open family. The 2.4T model’s licence requires products above 100M monthly active users or $20M monthly revenue to display the model name. Model-as-a-service businesses above $50M annual revenue need a separate licence.
Price and access
| Model | Input per million | Output per million | Weights |
|---|---|---|---|
| Qwen3.8-Max | $2 | $6 | Closed |
| Qwen3.8-Flash | $0.15 | $0.47 | Closed |
| Qwen3.8-27B | $0.50 | $3.00 | Apache 2.0 |
Prices are Alibaba Cloud’s international rates in US dollars. Alibaba lists a free quota of 1M tokens per model for 90 days.
Who should use it
On this evidence Qwen3.8-Max is a good choice for image-heavy products and web page generation. For agent pipelines, DeepSeek-V4.1-Flash, GPT-6.1 Sol or Qwen’s own Flash-Next give more per dollar. Developers who want to run a capable vision model locally should look at Qwen3.8-27B under Apache 2.0.
Sources
- qwen3.8-max Model Info, read 2026-09-30
- Alibaba Cloud Model Studio: Model inference pricing, read 2026-09-30
- Model releases - QwenCloud, read 2026-09-30
- Qwen3.8-2.4T-A95B on Hugging Face, read 2026-09-30
- Qwen3.8-2.4T-A95B licence, read 2026-09-30
- Qwen3.8-27B on Hugging Face, read 2026-09-30
- Qwen3.8-Flash-Next on Hugging Face, read 2026-09-30
- GitHub - QwenLM/Qwen3.8, read 2026-09-30
- Security certifications and privacy (Model Studio), read 2026-09-30
- LLM Leaderboard (Artificial Analysis), read 2026-09-30
- Top Open Source AI Models (Artificial Analysis), read 2026-09-30
- Arena vision leaderboard, read 2026-09-30
- Arena WebDev leaderboard, read 2026-09-30
- Vals Index, read 2026-09-30