Original caption
Inception calls it the most capable diffusion LLM released to date (October 19, 2026). Here's what you need to know. 1. It doesn't type. Most LLMs write one token at a time. Mercury generates and refines tokens in parallel, which is where the speed comes from. 2. Speed: 1,068 tokens per second on NVIDIA GPUs in Inception's own table, against about 240 for Claude Haiku 5.5 and about 130 for GPT-6 Luna. 3. Quality: it leads 8 of the 12 capability benchmarks in that table, including IFBench 76.84, SWE-Bench Atlas 51.08 and τ³-bench Banking 30.24. 4. Diffusion scales: more than 40x the parameters of Mercury 1, and about 3x the intelligence score from Mercury 1 (mini). 5. Price and access: $0.10 in and $0.35 out per million tokens at launch (50% off $0.20 / $0.70), cached input 90% off, 1M context. On Inception's API with 100M free tokens, OpenRouter, Baseten and Vercel, and it's OpenAI API compatible. The catch: every number is Inception's own, run at medium settings. It trails on AA-Omniscience and AA-LCR (GPT-6 Luna) and on MRCR v2 and DeepSWE (Claude Haiku 5.5), and OpenRouter showed a 383 tokens/s median on launch day, well under the lab figure. Follow @codenameposhan for more AI breakdowns. Sources: Inception's launch post (Oct 19, 2026) and the OpenRouter Mercury 3 page. Independent briefing by CodenamePoshan, not an Inception post. Inception's logo and launch art belong to Inception, shown for editorial reporting. #Inception #Mercury3 #DiffusionLLM #LLM #AI #MachineLearning