Daily AI CatchupInception claims Mercury 3 hits 1,000 tokens per second
Inception says its new Mercury 3 model generates over 1,000 tokens per second, using a diffusion approach instead of the token-by-token generation most chatbots use. The company is targeting interactive applications where latency matters, like voice agents and live coding assistants. The speed claim is Inception's own and hasn't been independently verified.
Why it matters
💻 Developer · If the speed claim holds up under real benchmarks, diffusion-based LLMs could change latency budgets for voice agents and live autocomplete tools.
📦 Product · Faster generation means snappier AI features with less need for workarounds like streaming tricks or response caching.
🎨 Design · Lower latency could finally make voice and conversational UI feel truly real-time instead of the stilted pause-and-respond pattern users tolerate now.
📈 Business · Speed is becoming a competitive axis alongside accuracy and price, worth tracking if you're evaluating model vendors for latency-sensitive products.
🤔 Just Curious · Most chatbots write one word at a time. This one claims to write whole chunks at once, like a sculptor refining a block rather than typing a sentence.
Sources: Inception says Mercury 3 generates over 1,000 tokens a second