Inception has released Mercury 3, a diffusion-based language model that generates and refines multiple parts of an answer in parallel rather than producing text strictly one token at a time. The design targets applications where repeated model calls make latency especially costly.
The company reports throughput of 1,068 tokens per second on Nvidia GPUs and says Mercury 3 beat GPT-6 Luna and Claude Haiku 5.5 in eight of 12 internal tests. Results were mixed: Mercury led instruction-following and one autonomous-coding test, while Claude won a longer coding benchmark and long-document retrieval. No independent evaluation is available yet.
Mercury 3 accepts contexts up to one million tokens and exposes an OpenAI-compatible API. It is available through Inception, OpenRouter, Baseten and Vercel. Early partners reported faster smart-glasses search, tutoring feedback and customer-service responses.
Regular pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a temporary 50% launch discount. Inception has not disclosed when that discount ends or the model’s exact parameter count, so developers should test quality, latency and eventual cost on their own workloads.