Alibaba’s Qwen team released Qwen3.8-Omni-Flash on September 18, a model built to read text, images, audio and video in a single pass and act as the sensory layer for autonomous agents. The launch lands as Chinese labs race to match Google’s Gemini line on multimodal reasoning while undercutting it on price. Qwen’s own numbers make a specific, and only partially supported, case for that trade.

The team says the model handles a million-token context window and stays nearly as strong on text tasks as a text-only model of equal size would be. Across 29 internal evaluations, Qwen reports an average score improvement of more than 25 percent over its predecessor, Qwen3.5-Omni-Plus, with gains of 36.5 points on WildClawBench-MM and 22.3 points on AgenticVBench, plus a 69.6 score on UniClawBench.

Pricing is where the announcement gets aggressive. Qwen says API costs fall by more than 98 percent for an hour of audio input and by more than 93 percent for an hour of combined audio and video. That is a cut steep enough to change the math for any team running always-on meeting transcription or customer-service agents at scale.

The comparison to Google’s Gemini 3.8 Flash is the claim to read carefully. Qwen’s team describes its own audio-visual performance as “close to” Gemini’s and its overall audio performance as exceeding it. The company’s own benchmark table tells a more mixed story: Gemini scores higher on OmniVideoBench (65.2 versus 63.4) and on Video-MME-v2 (71.0 versus 65.0). On multi-speaker transcription, Qwen posted AliMeeting error scores of 3.4 and 17.2, far below Gemini’s 72.6 and 53.1. Alibaba is ahead on audio, behind on several video benchmarks, and the “close to” language in its own release understates that gap.

The more interesting engineering claim is agentic video handling. Instead of scanning every frame, the system chooses which moments to watch and which sounds to register, a selection process Qwen says lifted OmniVideoBench accuracy from 63.4 to 67.8 while cutting tokens per query from 145,736 to 79,117, a drop of roughly 45.7 percent. That efficiency gain, if it holds outside Qwen’s own tests, matters more for agent economics than the headline benchmark score does. Token cost scales directly with how much of a video an agent has to ingest to answer a question.

Qwen also reported a self-improvement test in which the model was given 12 hours to raise a smaller model’s accuracy on Sichuan dialect speech recognition. Across four rounds, the model produced 3,413 training examples for a smaller sibling, Qwen2.5-Omni-3B. Its character-error rate dropped to 15.30 percent from a prior 25.79 percent, a cut of about 40.7 percent. The team is open-sourcing Qwen-MM-Plugins, tool integrations for Codex, Claude Code, Qwen Code and Gemini CLI, alongside a separate real-time variant built for live audio-visual streams such as speaking practice and customer service.

Every figure in this release, from the benchmark deltas to the price cuts, is Alibaba’s own measurement, not an independent evaluation. Teams building meeting-transcription or video-agent products on Gemini’s API should treat Qwen3.8-Omni-Flash as a price benchmark to take back to Google, not yet as a drop-in replacement, given Gemini’s own lead on the video tasks Qwen chose to publish.

Based on Alibaba’s Qwen team’s official blog post announcing Qwen3.8-Omni-Flash, published September 18, 2026.