Mercury 3 [Inception Labs] - Matching Luna/Haiku at 1000 tokens/sec

Explain Like I'm Normal

Inception Labs has released Mercury 3, a diffusion-based language model that claims to match the performance of Haiku and Luna while hitting extreme speeds of 1000 tokens per second. Unlike standard auto-regressive transformers, this architecture uses diffusion to generate text, potentially solving the sequential bottleneck that slows down current LLMs. If the benchmarks hold up, this could be a massive shift for high-speed agentic workflows.