Inception has released Mercury 3, a new diffusion language model achieving 1,068 tokens per second on NVIDIA GPUs with a 1 million token context window. Unlike traditional autoregressive models that generate text one token at a time, Mercury 3 utilizes a diffusion approach to write answers in parallel by starting from a rough draft of the full response and refining all positions over a few passes.
According to Inception's benchmark table, Mercury 3 leads rival models including GPT-6 Luna and Claude Haiku 5.5 on 8 out of 12 tests, with notable gains in instruction following and agentic coding, such as an IFBench score of 76.84 and a SWE-Bench Atlas score of 51.08. However, the model trails competitors in long-context retrieval and certain longer coding tasks like DeepSWE v1. The model is currently available via the Inception API Platform, OpenRouter, Baseten, and Vercel with an OpenAI-compatible API, alongside two sibling models named Mercury Voice and Mercury Decide.
This high-speed diffusion architecture matters for developers building real-time agents, voice bots, and search tools that require rapid latency and large context handling.