Answers questions about images, hour-long videos and long documents, with a dial for how hard it thinks

You send it text, images, or video, and it answers, extracts, or writes code from what it sees. Video is the part that stands out: it accepts an hour-long recording as input and answers questions about what happened across the whole thing, not just a few sampled frames. It also reads about 260,000 tokens in one prompt, a few books' worth.

Every answer starts in thinking mode, and you set the effort: low for a quick caption, high for a chart that needs careful reading. Qwen reports it ahead of its own Qwen3.8-27B on long video understanding, real-world photo questions and chart analysis, and ahead of DeepSeek V4 Flash on most coding and agent tests.

It holds 125 billion parameters but uses only 6 billion for each word, which is why it responds like a much smaller model. The weights are free to use commercially unless you serve more than 100 million monthly users or sell model access itself.

There is no official browser demo, and the weights are far too big for a desktop card, so the honest first test is the hosted version: Qwen Cloud serves it as Qwen3.8-Flash through an OpenAI-compatible API. Send a photo of a whiteboard and ask it to turn the diagram into a checklist, then repeat the request with reasoning effort set to low and see how much the answer changes. A community-run free endpoint on Hugging Face works without an account, with rate limits. To feed it video, serve the weights yourself with vLLM or SGLang; the model card names the one config change that unlocks hour-scale clips.

It is labeled experimental, and the scores are Qwen's own. Long-video answers depend on how many frames the server samples: with default settings it looks at a clip coarsely, and the setting for hour-long video needs far more memory.

• Summarize an hour-long lecture or meeting recording, then ask follow-up questions about specific parts.
• Turn a photo of a whiteboard or a hand-drawn flow into a written plan.
• Pull the numbers out of a chart in a paper and ask what they imply.
• Drive a phone or desktop agent from screenshots; it posts the top score on Android tasks in Qwen's table.

View model

Read this issue on neuralexpedition.com →