Inception unveils mercury 3, a high-speed diffusion LLM with 1,068 tokens per second

Inception has launched Mercury 3, its latest diffusion language model, capable of processing 1,068 tokens per second on NVIDIA GPUs. The model features a 1M-token context window and is designed for speed, outperforming GPT-6 Luna and Claude Haiku 5.5 on 8 out of 12 benchmarks. Mercury 3 is the fourth iteration in Inception’s Mercury series, following Mercury 1, Mercury 2, and Mercury 2.5. It boasts over 40 times the parameters of Mercury 1 and shows a 3x improvement in average quality across standard benchmarks. The model excels in instruction following and agentic coding, achieving scores of 76.84 on IFBench and 51.08 on SWE-Bench Atlas. However, it lags in long-context retrieval and extended coding tasks. Mercury 3 is available via Inception’s API, OpenRouter, Baseten, and Vercel, with pricing at $0.10 and $0.35 per 1M input and output tokens during its launch discount. Inception also released two additional models alongside Mercury 3.