ElevenLabs launches Eleven Realtime voice models with Mini and Max versions

ElevenLabs introduced Eleven Realtime on October 20, offering two speech-to-speech models for voice agents. The company positions v1 Mini for everyday conversations and v1 Max for tasks requiring more complex reasoning.

ElevenLabs said Max began responding after an average of 0.57 seconds and scored 98.8% on Big Bench Audio, a test containing 1,000 spoken reasoning questions. Mini recorded 0.62 seconds and 75.9%, respectively. The company attributed those results to independent testing.

The launch combines components that developers otherwise configure separately, including speech recognition, conversation timing and voice generation. Mini targets assistants, tutoring and interactive characters, while Max is intended for support calls and workflows involving several steps.

The models are available through ElevenAgents and the company's Realtime API. ElevenLabs says its API supports existing OpenAI Realtime integrations by changing the destination endpoint and selected model. Both versions support over 90 languages.