← Back to feed

Alibaba’s Qwen Team has launched Qwen3.8-Omni-Flash, a next-generation native omnimodal model designed to push audio and video from perceptual inputs into active agentic tools. The model accepts text, image, audio, and video in a single 1M-token context window, and it ships today on the Qianwen AI Platform.

What’s New

Qwen3.8-Omni-Flash is not a refinement of the text-focused Qwen3.8 Flash line. It targets workflows the prior generation couldn’t sustain: video editing, music video creation, film commentary, audio-visual summarisation, and real-time conversation agents that reason over multimedia streams.

Across 29 evaluations, the model’s average score improves by more than 25% over Qwen3.5-Omni-Plus. On the benchmarks that measure combined audio-visual agent capability:

• WildClawBench-MM (audio-visual agents, coding, long-horizon tasks): +36.5 points
• AgenticVBench: +22.3 points
• UniClawBench: 69.6

Audio-specific gains are meaningful: +8.3 points on LongAudioSpan, +9.6 points on OmniVideoBench. Multi-speaker recognition showed the sharpest jump: AliMeeting DER fell from 88.11 to 3.35 (word error rate cpWER from 89.61 to 17.18).

Against Gemini

Alibaba benchmarks audio-visual performance as close to Gemini 3.8 Flash, with overall audio performance exceeding it. The comparison is on model-level evaluations, not system-level integrations.

Pricing

The pricing change is the clearest signal that Alibaba is treating omnimodal as a commodity layer rather than a premium:

• API price per hour of audio input: down more than 98% vs prior generation
• API price per hour of audio-visual input: down more than 93%

That makes sustained audio-visual agent workflows — previously prohibitively expensive — practically viable at scale.

Text Performance

The model preserves text-only benchmark parity with a text-only model of the same parameter count. Maintaining quality across modalities without text regression has historically been the central engineering challenge in omnimodal training.

Context

Qwen3.8-Omni-Flash arrives in the same week that Qwen3.8 27B received a low-VRAM quantization from Byteshape, putting the 27B dense model into 13.1 GB. The Omni-Flash targets a different constraint: not VRAM on a single GPU, but multimodal agent pipelines that need to process long audio and video streams without chunking.

The model is available via API at Qianwen AI Platform. Weights availability was not announced.