Alibaba's Qwen Audio 3.0 Dethrones OpenAI With a 99.2% Reasoning Score
Alibaba's Qwen Audio 3.0 Realtime Plus tops every benchmark on the Artificial Analysis Speech-to-Speech Index, beating GPT-Realtime-2.1 High on intelligence at a fraction of the cost β but with a significant latency trade-off.
- New #1: Qwen Audio 3.0 Realtime Plus tops the Artificial Analysis Speech-to-Speech Index at 84.1%, beating GPT-Realtime-2.1 High (79.1%).
- Clean sweep: Plus variant leads all three component benchmarks β Big Bench Audio (99.2%), Full Duplex Bench (98.4%), and π-Voice (54.6%).
- Latency trade-off: Average Time to First Audio is 4.02s for Plus and 4.16s for Flash β vs. 1.14s for GPT-Realtime-2 High and 0.44s for the fastest model (Deepslate Opal).
- Pricing: Plus costs $4.42/hr of input audio β cheaper than GPT-Realtime-2.1 High ($10.75) but slightly more than GPT-Realtime-2 High ($4.14).
- Architecture: Built on Qwen3.5-Omni's Thinker-Talker MoE design with ARIA alignment for stable streaming speech and a custom AuT audio encoder trained on 40M hours of audio.
- Access: Available now via Alibaba Cloud DashScope API (WebSocket/WebRTC); new international accounts get 1M free tokens for 90 days.
Alibaba's Qwen team just claimed the top spot on the Artificial Analysis Speech-to-Speech Index with Qwen Audio 3.0 Realtime Plus, scoring 84.1% overall , nearly 5 points ahead of the previous leader, GPT-Realtime-2.1 High at 79.1%. It's the first time a non-OpenAI model has led this leaderboard, and the margin isn't trivial. The model comes in two variants: Plus (optimized for intelligence) and Flash (optimized for speed), both available now via Alibaba Cloud's DashScope API.
A clean sweep across all three benchmarks
The Artificial Analysis index is a composite of three separate evaluations, each testing a different dimension of voice AI capability. Qwen Audio 3.0 Realtime Plus leads all three , not just the overall score.
- Speech Reasoning (Big Bench Audio): Qwen Audio 3.0 Realtime Plus leads with a speech reasoning score of 99.2% across 32 models evaluated. This benchmark tests logical reasoning delivered as audio , things like counting objects, evaluating boolean logic, and detecting formal fallacies. The Flash variant scores 96.1%.
- Conversational Dynamics (Full Duplex Bench): Qwen Audio 3.0 Realtime Plus leads with a conversational dynamics score of 98.4% on the Full Duplex Bench dataset. This measures how well the model handles real conversation flow , knowing when to speak, when to pause, how to handle interruptions, and how to respond to backchannels like "mm-hmm." Flash follows at 96.9%.
- Agentic Performance (π-Voice): The Plus variant scores 54.6% on π-Voice, a benchmark that tests task completion in simulated customer service scenarios , things like changing a flight, disputing a retail charge, or resolving a telecom issue. It edges out Grok Voice Think Fast 1.0 at 52.1%, with Flash scoring 35.9%.
The architecture behind it
Qwen Audio 3.0 Realtime is built on the Qwen3.5-Omni foundation, which uses a Thinker-Talker architecture , a design pattern Alibaba introduced in Qwen2.5-Omni. The idea is conceptually clean: the Thinker is a large language model that handles reasoning, comprehension, and text generation, while the Talker is a streaming speech generation module that converts semantic representations and tokenized text into realistic audio.
In Qwen3.5-Omni (the underlying model), both components were upgraded significantly. Architecturally, Qwen3.5-Omni employs a Hybrid Attention Mixture-of-Experts (MoE) framework for both Thinker and Talker, enabling efficient long-sequence inference. The audio encoder (called AuT) was trained from scratch on 40 million hours of audio data, replacing the Whisper encoder used in earlier versions.