Mercury 3

Mercury 3 is Inception's flagship diffusion LLM, released 2026-10-19 (OpenRouter lists 2026-10-16). Rather than generating token by token it drafts a full answer and refines it over several passes; Inception claims ~1,068 tokens/s on NVIDIA GPUs (OpenRouter measured a ~200 tok/s weekly median). Parameter count undisclosed (Inception says >40x larger than Mercury 1 at the same speed). Text in / text out, 1M-token context (up from 260K), 131,072 max output; tool calling, structured outputs and four reasoning-effort levels; OpenAI-compatible API (model id mercury-3). List price $0.20 / $0.70 per 1M input/output ($0.02 cached), with a 50%-off launch price. Closed weights. Replaces Mercury 2 / 2.5, which are closed to new users with existing access ending 2026-11-15. Vendor benchmarks (medium effort): beats GPT-6 Luna and Claude Haiku 5.5 on 10 of 12 benchmarks each; GPQA Diamond 89.23, SWE-Bench Atlas 51.08.