Mercury 3 Claims 1,000 Tokens/Sec: Is Diffusion the Future of AI?

According to the-decoder.com, startup Inception has released Mercury 3, a language model built on diffusion technology. Instead of processing text word by word like standard models, it sketches a rough draft of the entire response and refines it across multiple passes.

How fast is Mercury 3 in numbers?

Inception claims the model generates more than 1,000 tokens per second on Nvidia GPUs. In internal company benchmarks cited by the-decoder.com, Mercury 3 runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5, outperforming both in eight out of 12 tests.

How does diffusion change real-time tasks?

Standard language models create a bottleneck because each generated word must wait for the preceding one. By generating many words at once, diffusion models reduce latency significantly for time-sensitive tools such as smart glasses and autonomous AI agents.

How to access and test Mercury 3

Developers can integrate Mercury 3 immediately using the same API structure as OpenAI. The model is currently available through Inception's own platform as well as third-party providers like OpenRouter, Baseten, and Vercel.

Is it worth using for your projects?

Mercury 3 is ideal for use cases where traditional response delays disrupt user experience, such as voice assistants, real-time search, and customer support. However, users should note that all performance claims currently rely entirely on the-decoder.com reporting of Inception's internal benchmarks, with no independent validations published yet.