Inception Labs highlights Mercury 2, the world's first reasoning diffusion LLM capable of sub-300ms reasoning passes at over 1,000 tokens per second for real-time voice agents.
Inception Labs has highlighted Mercury 2, a high-performance diffusion-based large language model designed to power real-time voice agents and agentic workflows. By replacing traditional sequential token decoding with a parallel refinement architecture, Mercury 2 completes a full reasoning pass in under 300 milliseconds while generating output at speeds exceeding 1,000 tokens per second on standard NVIDIA GPUs. This model effectively resolves the trade-off between reasoning quality and latency, enabling production-grade voice agents capable of handling live phone calls.
Mercury 2 demonstrates that diffusion architectures, not traditional sequential decoding, are the key to unlocking the real-time reasoning speeds required for agentic workflows.
* The parallel refinement process bypasses the token-by-token generation bottleneck of standard LLMs.
* Sub-300ms latency makes deep reasoning viable for phone calls and voice agents where human-like response speed is critical.
* Optimizing for standard NVIDIA GPUs ensures cost-effective scalability for enterprise contact centers.
* By targeting the intersection of reasoning and real-time voice, Inception Labs is addressing a major bottleneck in agent adoption.
DISCOVERED
1h ago
2026-07-14
PUBLISHED
1h ago
2026-07-14
RELEVANCE
AUTHOR
_inception_ai
