Alibaba's Qwen team recently released the new native multimodal model Qwen3.8-Omni-Flash, further integrating audio and video understanding with AI agent capabilities. It supports four types of input: text, images, audio, and video, and provides a 1 million token context window. Compared to the previous generation Qwen3.5-Omni-Plus, the official claims an average improvement of over 25% in 29 benchmark tests.
The biggest change this time is not about adding features, but about the processing method: instead of simply combining speech and image recognition, it natively handles different types of information at the model level. Alibaba's intention is clear — to make the model not only "understand" audio and video, but also plan tasks, call tools, and complete work after understanding.
Audio is its strongest aspect. In the WildClawBench-MM multimodal tool calling test, Qwen3.8-Omni-Flash scored 71.0, while Gemini3.8Flash scored 58.9; on DailyOmni, it was 85.1 versus 84.0; on SpotSoundBench, 67.2 versus 39.7; and on MMAU, 81.8 versus 76.9, winning all four. The improvement in multi-speaker meeting recognition is even more impressive: in the AliMeeting test, the speaker error rate (DER) dropped from 88.11 in the previous generation to 3.35, and the word error rate with speaker attribution (cpWER) fell from 89.61 to 17.18. However, it is not leading in all areas — on AgenticVBench, Gemini3.8Flash scored 45.0, while Qwen scored 36.8; on OmniGAIA, it was 78.6 versus 74.0, and in some video understanding tests, Gemini still had an advantage. Therefore, the claim that it is "approaching Gemini" mainly applies to overall audio and audio-video capabilities.
The real change that may rewrite usage patterns is the price. Qwen claims that the estimated cost for audio input has dropped by more than 98% per hour, and for combined audio and video input, it has dropped by more than 93%. According to the official statement, the hourly cost is calculated as 30 times the price of processing two minutes of material, with video input at 720p and one frame per second. Alibaba Cloud's international region pricing is $0.15 per million input tokens, $0.016 per million cached input tokens, and $0.47 per million output tokens, with different prices for mainland China and some other regions.
With a 1 million token context window (maximum input close to 992,000, around 984,000 in thinking mode, and maximum output of 131,000), developers can directly feed extremely large amounts of audio or video to the model without frequently slicing, extracting frames, or connecting multiple models. The model can also handle audio input in 113 languages and dialects, support stereo and 4-channel spatial audio analysis, and provide function calls, online searches, and context caching.
In terms of application directions, the Qwen team has focused on "agent delivery": the model can analyze long videos, locate key content, and then use tools to perform video editing, content organization, meeting summaries, and subtitle generation. Accompanying tools include Qwen-MM-Plugins for multimodal agent development, and they plan to launch Qwen-Live-Harness. Currently, Qwen3.8-Omni-Flash is available through Alibaba Cloud BaiLian, covering regions including Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia. The model itself is not directly open-sourced, but this release mainly opens up the accompanying multimodal plugins and development tools.
If the more than 98% reduction in audio costs is realized, the automatic analysis of long meetings, podcasts, live streams, and massive video materials will become exceptionally cheap. The closely fought competition between Qwen and Google Gemini in the multimodal field will thus escalate once again.
