- Published on
Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understandin...
- Authors
- Name
- DeepHorus
Alibaba’s Qwen team has released Qwen3.8-Omni-Flash, its first omni-modal model built around agentic capabilities. The model accepts text, images, audio, and video, and returns text, with audio-video understanding, reasoning, and tool use integrated into a single system. It is available today as a hosted API on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio. No open weights were announced at launch, so self-hosting is not currently an option.
Qwen3.8-Omni-Flash is built on the Qwen3.8-Flash-Next architecture, which shipped with open weights in August 2026. The context window is 1M tokens, with QwenCloud listing 991K max input and 131K max output, plus a max reasoning length of 262K tokens. Output is text-only; developers needing generated speech are directed to Qwen3.5-Omni. Thinking is enabled by default with reasoning_effort set to xhigh, and can be disabled by setting it to none.
The API follows both DashScope and OpenAI protocols, supporting Chat Completions and the Responses API. Function calling, web search, structured outputs, context caching, and batch calls are all supported. The model’s agentic perception differs from typical video models: it starts from the question, decides what to watch and hear, and gathers evidence over several coarse-to-fine rounds. On OmniVideoBench, accuracy rose from 63.4 to 67.8 while token use dropped from 145,736 to 79,117, a reduction of about 45.7%.
Across 29 evaluations, Qwen reports an average score improvement of more than 25% over Qwen3.5-Omni-Plus. WildClawBench-MM improved by 36.5 points, AgenticVBench by 22.3 points, and UniClawBench reached 69.6. LongAudioSpan gained 8.3 points, OmniVideoBench gained 9.6 points, and OmniCap-IF CSR and ISR improved by 8.5 and 14.1 points respectively. The research team claims audio-visual performance is close to Gemini 3.8 Flash, with overall audio performance above it. Independent results were not available at publication.
Pricing is set at $0.15 per 1M input tokens and $0.47 per 1M output tokens, with implicit cache hits costing $0.016 per 1M tokens. Qwen reports large cost cuts against Qwen3.5-Omni-Plus: audio input costs over 98% less per hour, audio-visual input over 93% less, and video input about 89% less. This aggressive pricing could make long-form multimodal analysis far more accessible to developers.
Model Studio docs list video files up to 2 hours and 2GB by URL, audio files up to 3 hours, and audio input in 113 languages and dialects. Stable results are expected with video sampled at up to 15 fps, and two-channel stereo and four-channel FOA spatial audio are supported through use_multichannel. Availability spans six regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia.
The research team frames the release as a shift toward question-driven, tool-using multimodal agents rather than brute-force full-file processing. The X post summarizes agent gains as +19.5 points on average across WildClawBench-MM and UniClawBench. Open-source tooling, including Qwen-MM-Plugins and Qwen-Live Harness, accompanies the model, though details were not fully expanded in the announcement.
The hosted-only launch suggests Alibaba is prioritizing commercial deployment and developer feedback before any open-weight release. With a 1M context window and aggressive pricing, the model positions Qwen to compete directly with frontier omni-modal systems. Future iterations may expand output modalities or open-weight availability, but no timeline was given.
