Qwen2.5-Omni

// Overview

Qwen2.5-Omni is Alibaba's (Qwen) first end-to-end omni-modal model line, released on March 26, 2025. A single model takes text, images, audio and video as input and generates both text and natural speech, streaming the speech back in real time. It shipped as open weights in two sizes, a 7B and a smaller 3B variant, under the Apache-2.0 license.

The model uses a Thinker-Talker architecture: the Thinker is a transformer-decoder LLM (with audio and vision encoders) that understands the inputs and produces text, while the Talker is a dual-track autoregressive decoder that turns the Thinker's representations into discrete speech tokens. To keep video frames and audio in sync, Qwen2.5-Omni introduces TMRoPE (Time-aligned Multimodal RoPE), a position embedding that aligns the timestamps of video and audio. The 7B model supports a 32,768-token context and ships with two built-in speech voices, Chelsie (female) and Ethan (male).

Qwen2.5-Omni reports state-of-the-art results on OmniBench, the cross-modal reasoning benchmark, and its performance following spoken instructions is close to its performance on the same tasks given as text. Besides the open weights on Hugging Face and GitHub, it is served through Alibaba Cloud Model Studio (DashScope) as qwen2.5-omni-7b. It was succeeded by Qwen3-Omni in September 2025.

// Benchmarks

1. OmniBench (avg)56.13%
2. MMLU-redux71%
3. MMMU (val)59.2%
4. MMStar64%
5. MVBench70.3%
6. GSM8K88.7%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Alibaba Cloud Model Studio (Singapore), qwen2.5-omni-7b, text input / text-only output; multimodal-input output is $0.84/1M.

Pricing source ↗

// Strengths

• Truly omni-modal: one model accepts text, image, audio and video and produces both text and speech, without stitching separate models together.
• Real-time, streaming speech output with two natural-sounding built-in voices (Chelsie and Ethan).
• Open weights under Apache-2.0 in both 7B and 3B sizes, so it can be self-hosted and used commercially.
• Strong cross-modal reasoning: state-of-the-art OmniBench average (56.13) with spoken-instruction following close to text-instruction quality.
• TMRoPE keeps audio and video frames time-aligned, improving understanding of synced audio-visual content.

// Best for

• Real-time voice assistants and conversational agents that listen and talk back.
• Audio-visual understanding: answering questions about video clips with synced sound.
• Speech-to-speech and spoken-instruction following without a separate ASR/TTS pipeline.
• On-device or self-hosted multimodal apps using the smaller 3B variant.
• Building accessibility and voice-interface features on open, commercially-licensed weights.

// How to access

// Qwen-Omni — every version

The full lineage of the Qwen-Omni line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.