Inception details Mercury 2 voice capabilities

By achieving sub-300ms reasoning latency, Mercury 2 breaks the primary barrier preventing reasoning models from powering fluid voice interfaces. Parallel token generation through diffusion represents a major architectural paradigm shift that could make traditional sequential autoregressive models obsolete for interactive, multi-step agentic workflows.

• –A 300ms reasoning pass finally matches the natural cadence of human speech, making real-time interactive voice agents viable.
• –Moving away from autoregressive decoding to parallel diffusion refinement allows for extreme throughput (>1,000 tok/s) on standard GPUs.
• –Beating GPT-4.1 by 25 points on IFBench and 11 points on Tau3Bench shows that speed does not come at the cost of instruction-following capability.
• –At ~0.5¢ per conversation minute, it drastically lowers the operating cost of agentic voice and customer support systems.