New Qwen Models on Model Studio: qwen3.8-omni-flash and More
Between September and early October 2026, Alibaba Cloud Model Studio (known in China as Bailian) added a batch of new Qwen models: the omni-modal qwen3.8-omni-flash, realtime voice and realtime audio/video models, a decision model called decision-model-preview, and two image models, qwen-image-2.1-pro and qwen-image-2.1-turbo. This post follows the official "Newly released models" page and reflects its state as of October 17, 2026. If the page has changed since, trust the official version. The page does not list prices for these entries, so this post does not either.
Which models did Model Studio add in September and October?
The source is the Model Studio newly released models page (Chinese). The Qwen-related entries are:
| Date | Model | Highlights as listed on the page |
|---|---|---|
| Sep 17 | qwen3.8-omni-flash | Omni-modal; thinking and non-thinking modes; function calling; web search; context cache |
| Sep 20 | qwen-audio-3.1-realtime-plus | Realtime full-duplex voice conversation; function calling; web search; voice cloning |
| Sep 21 | qwen3.8-omni-flash-realtime | Realtime omni-modal; WebSocket and WebRTC access; remote MCP tool calls |
| Sep 24 | decision-model-preview | Decision model; parallel classification, yes/no judgments and scoring, returns a probability distribution and confidence |
| Oct 2 | qwen-image-2.1-pro | Text-to-image and image editing in one model; transparent (RGBA) output; up to 10 reference images |
| Oct 9 | qwen-image-2.1-turbo | Listed as an open-source image generation model balancing performance and cost |
The page also records several launches in the US (Virginia) region: qwen3.8-flash on October 3, and qwen3.8-max and qwen3.7-flash on October 8. These three are not new October releases. What changed is that they became selectable in the US region. For example, qwen3.7-flash has a July 21 entry on the Germany and Japan pages. On September 17, qwen3.8-omni-flash appeared on the Virginia, Frankfurt and Tokyo pages the same day (deployment scope is marked "global"), and it is also available in Beijing and Singapore.
For context length, output limits, regions and rate limits of each model, check that model's detail page.
Which new model fits which use case?
| What you are building | Start with | Notes |
|---|---|---|
| Handling text, image, audio and video input together | qwen3.8-omni-flash | The page lists thinking and non-thinking modes, so you can toggle per task |
| Voice assistants, phone-style realtime conversation | qwen-audio-3.1-realtime-plus | Listed as realtime voice dialogue, with voice cloning |
| Realtime audio/video interaction that needs external tools | qwen3.8-omni-flash-realtime | The page lists remote MCP tool calls |
| Ticket routing, content moderation, yes/no judgments | decision-model-preview | Returns a probability distribution and confidence; the name says preview, so evaluate before production |
| Generating or editing images | qwen-image-2.1-pro / turbo | See the model detail pages for the differences |
The decision model deserves a closer look. Many teams use a general-purpose LLM just to get a single label. If your workflow can use a confidence threshold (below it, hand off to a human or to a larger model), a dedicated model may be a better fit. But cost and quality have to be measured on your own data, so don't draw conclusions up front.
How do I call these with the OpenAI-compatible API?
Model Studio offers an OpenAI-compatible endpoint. Here is a minimal example. Check the official docs for the endpoint URL, region and key rules:
curl https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions \
-H "Authorization: Bearer $DASHSCOPE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-omni-flash",
"messages": [{"role": "user", "content": "Introduce yourself in one sentence"}]
}'
A few things to watch:
- Keys are tied to regions. Endpoints and keys for Beijing, Singapore, the US and other regions are not necessarily interchangeable. Confirm in the docs which region your model is in and which endpoint goes with it.
- Realtime models use WebSocket or WebRTC, not the plain HTTP request above. Follow each model's documentation.
- Models marked preview may change in name or behavior, so don't hardcode the model name somewhere that is hard to change.
How do I verify before going live?
Four steps, in order:
- On the cdnping home page, test first-connection and round-trip latency to the Alibaba Cloud Model Studio API. This measures the connection from your browser to the API entry point, not model inference time. For how to read the results, see how to measure latency to AI APIs from mainland China.
- Open /en/status/dashscope to see entry-point probes and 30-day history, and confirm the entry point is stable.
- Run 50 to 100 real samples of your own, record accuracy, latency and failure rate, and only then decide whether to replace your current model.
- Subscribe to Model Studio's changelog at /en/changes so you see deprecations and adjustments right away. For a method, see how to find out the moment an LLM API price changes.
How does this compare with other vendors' updates?
Other vendors shipped model updates around the same time, so it helps to read them side by side:
- DeepSeek V4.1-Flash-1017 is live: the Model Studio page also lists deepseek-v4.1-flash, so if you are comparing the two routes, read both.
- Anthropic's three new models in September and October: the approach to watching availability and latency after launch applies broadly.
- How much uptime differs when you switch OpenRouter providers for the same model: a model with the same name can behave differently across providers, so verify before migrating.
Sources
- Alibaba Cloud Model Studio: Newly released models (Chinese; accessed October 17, 2026)
