Qwen3-Omni-Flash
An omnimodal large language model released by Alibaba's Qwen team. It supports multimodal inputs including audio, images, and video, and produces both text and speech outputs. The model adopts a Mixture-of-Experts (MoE) architecture and features long-form audio and video understanding as well as native agentic capabilities. It is made available to external users via API.
Timeline (last 90 days)
Qwen3-Omni-Flash has not yet been open-sourced and is made available externally through an API interface.
Qwen3-Omni-Flash supports audio, images, and video on the input side, and can generate both text and speech modalities on the output side.
Official evaluation data shows that Qwen3-Omni-Flash's overall audio capability surpasses the same-tier Gemini Flash version and represents a notable improvement over the previous generation Qwen Omni Plus.
Qwen3-Omni-Flash has seen a significant reduction in API call pricing alongside its performance improvements.
Qwen3-Omni-Flash can understand 74 global languages and supports recognition of 39 Chinese dialects.
The official compared Qwen3-Omni-Flash against the previous generation Gemini series and models such as U-Spark open-sourced by Manta.
All Facts (6)
Qwen3-Omni-Flash has not yet been open-sourced and is made available externally through an API interface.
50%UnverifiedQwen3-Omni-Flash supports audio, images, and video on the input side, and can generate both text and speech modalities on the output side.
50%UnverifiedOfficial evaluation data shows that Qwen3-Omni-Flash's overall audio capability surpasses the same-tier Gemini Flash version and represents a notable improvement over the previous generation Qwen Omni Plus.
50%UnverifiedQwen3-Omni-Flash has seen a significant reduction in API call pricing alongside its performance improvements.
50%UnverifiedQwen3-Omni-Flash can understand 74 global languages and supports recognition of 39 Chinese dialects.
50%UnverifiedThe official compared Qwen3-Omni-Flash against the previous generation Gemini series and models such as U-Spark open-sourced by Manta.
50%