Chinese AI startup StepFun dropped Step-Audio-R1.1 this week, and it immediately grabbed the #1 position on the Artificial Analysis Speech Reasoning leaderboard. The 33-billion-parameter model posted 96.4% accuracy, edging out major closed-source competitors including xAI's Grok, Google's Gemini, and OpenAI's GPT-Realtime.
The model builds on the original Step-Audio-R1, which StepFun released in late November with a technical paper claiming to solve a weird problem in audio AI: longer reasoning chains actually degrading performance. Most audio models inherit their reasoning habits from text training, so they analyze transcripts instead of listening to actual acoustic features like pitch, timbre, and rhythm. StepFun calls this "textual surrogate reasoning" and addressed it with a training method called Modality-Grounded Reasoning Distillation.
R1.1 adds a "Dual-Brain Architecture" that separates high-level reasoning from speech generation. The idea is the model can think while speaking instead of choosing between intelligence and response speed. StepFun reports sub-second first-packet latency on the Realtime variant, though independent verification is pending. The weights are on Hugging Face, built on a Qwen2.5-32B backbone. Full voice API expected in February.
The Bottom Line: An open-source model taking the top benchmark spot over proprietary systems is notable, though real-world voice agent performance will tell whether the numbers translate.
QUICK FACTS
- 96.4% accuracy on Artificial Analysis Speech Reasoning (company-reported ranking)
- 33B parameters on Qwen2.5-32B backbone
- Sub-second first-packet latency claimed for Realtime variant
- Model weights: open source on Hugging Face and ModelScope
- Full voice API launch: February 2025
