Mercury
lastmod 2025-07-09created 2025-07-091 edits
Ultra-fast language models based on Diffusion
Features
- Mercury is a new commercial large language model (LLM) that uses a Diffusion approach
- The model is based on the Transformer architecture and is characterized by predicting multiple tokens in parallel
- Mercury Coder is the first set of diffusion LLMs, developed for writing code, and comes in two sizes: Mini and Small
- On an NVIDIA H100 GPU it achieves throughput of 1109 (Mini) and 737 (Small) tokens/sec, up to 10x faster than existing speed-focused models at the same quality
- In real-world benchmarks and developer evaluations such as Copilot Arena, it ranked 2nd in quality and fastest in speed; a public API and playground are also provided
Documentation
- [2506.17298] Mercury - Ultra-Fast Language Models Based on Diffusion
- https://arxiv.org/abs/2506.17298