Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18, an omni-modal model that accepts text, images, audio and video and decides for itself which parts of a file are worth examining. On OmniVideoBench, that selective approach lifted accuracy from 63.4 to 67.8 while cutting token consumption from 145,736 to 79,117 β roughly 45.7% fewer tokens for a better answer, according to Qwen's launch post.
Key takeaways
- Agentic perception raised Qwen3.8-Omni-Flash's OmniVideoBench accuracy from 63.4 to 67.8 while using about 45.7% fewer tokens than reading a video end to end.
- Across 29 evaluations the model averages more than 25% above Qwen3.5-Omni-Plus, with WildClawBench-MM at 71.0 (+36.5 points) and AliMeeting speaker-diarization error falling from 88.1% to 3.4%.
- Unlike most of Qwen's recent headline releases, the model ships API-only at $0.15 per million input tokens and $0.47 per million output tokens, with no weights to download.
What agentic perception changes
Conventional video models ingest a file from beginning to end even when the answer lives inside a three-minute stretch. Qwen's researchers describe the opposite order of operations: the model starts from the question, decides which segments to watch and listen to, then gathers evidence across several coarse-to-fine passes.
Compute lands on the footage that carries the answer instead of being spread evenly across an hour of video. The efficiency gain is the point β the accuracy gain comes with it rather than at its expense, which is unusual for a technique that reads less.
How the benchmark sheet reads
Qwen reports gains across 29 evaluations, averaging more than 25% over the previous multimodal flagship Qwen3.5-Omni-Plus. The largest jumps land in agentic territory: WildClawBench-MM reaches 71.0, up 36.5 points, with UniClawBench at 69.6 and AgenticVBench improving 22.3 points.
Core perception moved too, though Qwen reported most of it as point gains rather than absolute scores. OmniVideoBench sits at 63.4 in static mode before agentic perception is applied, a 9.6-point gain, with LongAudioSpan up 8.3 points on long-audio comprehension and OmniCap-IF up 8.5 and 14.1 points on caption content structure and instruction following. The sharpest move is multi-speaker Chinese meeting transcription: AliMeeting diarization error falls from 88.1% to 3.4% and transcription error from 89.6% to 17.2%. Qwen positions audio-visual performance as close to Gemini 3.8 Flash and overall audio performance as ahead of it. Every figure is self-reported; no independent evaluation was available at launch.
What it costs and where it runs
QwenCloud lists $0.15 per million input tokens and $0.47 per million output tokens, with implicit cache hits at $0.016 per million. Against Qwen3.5-Omni-Plus, Qwen claims hourly audio input costs fell more than 98% and audio-visual input more than 93%.
The context window is one million tokens β QwenCloud specifies 991K maximum input and 131K output, with reasoning length capped at 262K. Video files are accepted up to two hours and 2GB by URL, sampled at up to 15 frames per second, with two-channel stereo and four-channel spatial audio supported. Speech recognition covers 74 languages plus 39 Chinese dialects. Serving runs from six regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia.
Closed model, open plumbing
The model itself is API-only. It is built on the Qwen3.8-Flash-Next architecture, whose weights Alibaba Qwen did release in August 2026 β a contrast the team did not repeat here. Output is text only; developers who need generated speech are pointed back to Qwen3.5-Omni.
What did open is the surrounding tooling. Qwen-MM-Plugins ships under Apache-2.0 with a guided installer for Claude Code, Codex, Gemini CLI, Qwen Code and other harnesses, exposing each capability as a skill plus an optional MCP server. A companion Qwen-Live Harness was announced alongside it, though its repository was still returning a 404 shortly after launch.
Why it matters for agent builders
The plugin layer signals what Qwen thinks the bottleneck is. Most agent harnesses cannot feed audio to a main model natively, so audio still routes through the API even with the plugins installed. By open-sourcing the connective tissue while keeping the model closed, Alibaba is betting that adoption comes from making audio and video ordinary inputs for agents β a different posture from the open-weight releases that built Qwen's reputation. Whether developers accept an API-only omni-model from a lab known for downloadable ones is the open question.
FAQ
Is Qwen3.8-Omni-Flash open source?
No. The model is available only as a hosted API through QwenCloud, Alibaba Cloud Model Studio and Qwen Studio, with no weights released at launch. Its base architecture, Qwen3.8-Flash-Next, did ship with open weights in August 2026, but that does not make the omni model self-hostable.
Can it generate speech?
Not in the standard API. Qwen3.8-Omni-Flash accepts audio and video but returns text only, and Alibaba's documentation directs developers who need synthesized speech to Qwen3.5-Omni instead. A separate realtime variant was referenced at launch for live conversational use.
How does it compare to Gemini 3.8 Flash?
Qwen claims audio-visual performance close to Gemini 3.8 Flash and overall audio performance above it. Those comparisons come from Alibaba's own evaluation runs across 29 benchmarks, and no independent third-party testing had been published at the time of release.
