Mercury Voice
Mercury VoiceA diffusion LLM for real-time voice agents, combining fast reasoning, tool calling, and low latency for natural conversations.
Model details
Mercury Voice is Inception’s diffusion LLM purpose-built for real-time voice agents. It combines reasoning, tool calling, and instruction following with low latency, so your application can respond intelligently without awkward pauses.
Mercury Voice accepts text input and generates text output. It powers the language-model layer of a voice application, with speech-to-text and text-to-speech handled by separate components.
Key capabilities:
Low-latency responses: 320 ms median time to first answer token on production voice prompts, with 750 ms p95 latency.
Real-time reasoning: Choose low, medium, or high reasoning effort.
Tool calling: Connect conversations to external tools and workflows.
Instruction following: Follow long system prompts for complex conversational tasks.
128K-token context: Support lengthy conversation histories and reference material, with up to 50K output tokens.
Mercury Voice is available to enterprise customers through Inception’s OpenAI-compatible API. It fits into the LLM layer of voice frameworks such as LiveKit, Pipecat, Vapi, and Retell. It’s best suited to latency-sensitive conversational applications, including customer-service calls, automated ordering, and financial-services voice agents.
