Inception has launched Mercury 3, a diffusion-based large language model that the company says can generate up to 1,068 tokens per second on Nvidia GPUs and handle a 1-million-token context window.
According to Inception, Mercury 3 is four to eight times faster than comparable speed-optimized models and matches or outperforms GPT-6 Luna and Claude Haiku 5.5 on selected benchmarks. The company noted that these comparisons have not all been independently verified.
The model uses an iterative generation process rather than producing text one token at a time, allowing more parallel processing. Inception said its parameter count is more than 40 times larger than Mercury 1, while internal evaluations showed roughly threefold improvement in benchmark intelligence between Mercury 1 mini and Mercury 3.
Mercury 3 is available through Inception’s API Platform, OpenRouter, Baseten and Vercel, with support for OpenAI-compatible API requests. Standard pricing is $0.20 per 1 million input tokens and $0.70 per 1 million output tokens. A 50% launch discount reduces those prices to $0.10 and $0.35, respectively, while developers can also access 100 million free tokens for evaluation.
The company said early users are testing Mercury 3 for web agents, education, real-time search and customer support. In one evaluation involving 50 live Mind2Web tasks, the model completed 50% of them under Inception’s test conditions.
Source: EntArabi
