Mercury 2.5 Hits 1,178 Tokens/Sec

🔥 What happened Inception's Mercury 2.5 clocks roughly 1,178 output tokens per second on Artificial Analysis, ranking #2 of 146 models. It uses a diffusion-style approach instead of sequential autoregressive decoding. The Intelligence Index is still unknown. 💡 Why it matters At 1,000+ tok/s, a 500-token answer streams in about half a second. An 8-step agent loop generating 2,400 tokens drops from tens of seconds to a few — making reasoning budgets, best-of-N sampling, and inline verification suddenly practical. ⚡ Our take Speed only multiplies per-step accuracy: a fast, weak model just fails faster. With quality unverified, this number is an architecture signal, not a production argument.