Qwen3.8-Omni-Flash: 1M-Context Audio, Video and Agent Model
Alibaba's Qwen team has released Qwen3.8-Omni-Flash, an API-served model that accepts text, images, audio and video and returns text. The model provides a 1 million-token context window, function calling, adjustable reasoning effort, web search and context caching through Alibaba Cloud Model Studio.
International QwenCloud pricing lists $0.15 per million input tokens, $0.47 per million output tokens and $0.016 per million cached input tokens. The service supports a maximum input of 991,808 tokens in non-thinking mode, 983,616 tokens in thinking mode and up to 131,072 output tokens.
The release targets multimodal agent workflows in which a single model must interpret mixed media and then use tools. Alibaba lists availability through Model Studio in China (Beijing), Singapore, Hong Kong, Japan (Tokyo), Germany (Frankfurt) and the US (Virginia).
Qwen3.8-Omni-Flash specifications
| Specification | Qwen3.8-Omni-Flash |
|---|---|
| Inputs | Text, image, audio, video |
| Output | Text |
| Context window | 1,000,000 tokens |
| Maximum input, non-thinking | 991,808 tokens |
| Maximum input, thinking | 983,616 tokens |
| Maximum output | 131,072 tokens |
| Function calling | Supported |
| Reasoning | Enabled by default; adjustable effort |
| Web search | Supported through Responses API |
| Context caching | Implicit caching and Responses Session caching |
| Spatial audio | Multichannel input supported |
| International input price | $0.15 / 1M tokens |
| International cached-input price | $0.016 / 1M tokens |
| International output price | $0.47 / 1M tokens |
Pricing above is the current international QwenCloud schedule. Alibaba Cloud uses region-specific service endpoints and can price regions separately, so production cost estimates should use the target deployment region.
What the omni model adds
Qwen3.8-Omni-Flash combines media understanding with agent-oriented capabilities in one API model. An application can submit a mixture of text, images, audio and video, ask the model to reason over that material, and expose custom functions for the model to call as part of the workflow.
Alibaba also supports built-in web search through the Responses API. This makes the model relevant to systems that need to analyze a meeting recording, screen capture, video, document images or other mixed-media input before retrieving current information or invoking an application tool.
The model supports multichannel spatial audio. Alibaba's current documentation lists audio support across 113 languages and dialects and exposes a use_multichannel option for spatial-audio input.
The 1M context changes multimedia workflow design
A one-million-token context gives developers substantially more room for long transcripts, document collections and multimodal session history than the 64K context listed for the preceding Qwen3.5-Omni family in Alibaba's current model documentation.
The usable input ceiling is slightly below the headline context size because space is reserved for generation and, in thinking mode, reasoning. Applications should therefore plan against the published 991,808-token or 983,616-token input limits instead of treating one million tokens as the maximum payload size.
Long context also makes caching more important. QwenCloud lists cached input at $0.016 per million tokens, versus $0.15 for ordinary input. Workloads that repeatedly send a large stable prefix can reduce repeated input cost through the platform's caching mechanisms.
Tool calling extends the model into agent workflows
Function calling is available alongside text, image, audio and video input. That combination supports workflows such as extracting events from a long recording and writing them into another system, inspecting visual evidence before querying a database, or analyzing multimedia support material before invoking a troubleshooting tool.
The model can be called through Chat Completions or Responses. Alibaba's documentation also provides OpenAI-compatible API examples, which reduces migration work for applications already built around that request format.
Deployment and availability
Qwen3.8-Omni-Flash is currently an inference service in Alibaba Cloud Model Studio rather than a local checkpoint in the release material reviewed for this article. Developers choosing it for production should therefore evaluate regional endpoint availability, data-handling requirements, rate limits and media-upload behavior alongside model quality.
The international QwenCloud page currently lists a 2 million-token-per-minute limit and 30,000 requests per minute. Actual account and regional limits can vary with the service configuration.
For teams building private local-AI systems, the API delivery model is a material architecture difference from downloadable Qwen checkpoints. For cloud applications, the more relevant questions are multimodal accuracy, latency, cache hit rate, tool-call reliability and total token consumption on representative media.
How to evaluate it
Alibaba reports substantial gains over its earlier omni models and publishes benchmark comparisons for audio, video and agentic multimodal tasks. Those figures are vendor measurements. Independent launch-day evaluation remains limited, so production selection should use a representative workload test set.
A useful evaluation should measure transcription or understanding accuracy on the languages and acoustic conditions actually used, video reasoning on representative clip lengths, tool-call success rate, first-token latency, total response latency, context-cache hit rate and cost per completed workflow.
For agent systems, end-to-end task completion matters more than modality support alone. A model that understands a long video accurately can still perform poorly in production if function selection, argument generation or multi-step recovery is unreliable.
Bottom line
Qwen3.8-Omni-Flash combines a 1M-token context window, four input modalities, function calling, reasoning, web search and low published international API pricing in one model. Its strongest deployment fit is multimodal agent software that needs to process substantial audio/video context and then act through tools.
The launch provides enough API, context, regional and pricing detail for developers to begin workload-specific testing. Independent benchmark evidence is still limited at release, making application-level evaluation the appropriate basis for production adoption.