Mercury 2
lastmod 2026-02-26created 2026-02-261 edits
A diffusion-based ultra-fast reasoning LLM
About
- A language model that overcomes the speed limits of conventional sequential-decoding LLMs by using parallel generation based on Korean: 확산 모델 (diffusion)
- Achieves more than 5x faster responses with a parallel refinement architecture that generates and revises multiple tokens at once
- Optimized for real-time applications with 1,009 tokens/sec throughput, 128K context, JSON output, tool use, and more
- Proven efficient in latency-sensitive environments such as coding assistance, agent loops, voice interfaces, and search/RAG pipelines
- Fully compatible with the OpenAI API; can be integrated right away without changing existing infrastructure