Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second

Inception Labs has released Mercury 3, a diffusion-based LLM featuring a 1M-token context window and speeds of 1,068 tokens/second on NVIDIA GPUs. It reportedly outperforms GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks. Available via OpenRouter, Baseten, Vercel, and an OpenAI-compatible API, it costs $0.20/$0.70 per million tokens, with half-price launch pricing. Unlike autoregressive models, Mercury 3 utilizes a different generation paradigm, altering standard inference assumptions for developers and builders.