Inception unveils Mercury 3 diffusion language model

Inception, a California-based AI research company, has unveiled Mercury 3, its latest diffusion-based language model. The new model is designed to provide high-performance reasoning while maintaining extremely low latency, reportedly generating over 1,068 tokens per second on NVIDIA GPUs.

Unlike traditional autoregressive models that generate text sequentially, Mercury 3 utilizes a diffusion architecture that generates and refines multiple tokens in parallel. This approach allows the model to achieve speeds four to eight times faster than competitors such as GPT-6 Luna and Claude Haiku 5.5 on various benchmarks.

Mercury 3 features a context window of one million tokens, a significant increase from the 260,000 tokens supported by Mercury 2.5. The company noted that the model is more than 40 times larger than its predecessor, Mercury 1, demonstrating that diffusion-based models can scale in capability without sacrificing the speed advantages of their architecture. The model will be distributed via OpenAI-compatible APIs.