Mercury Coder

Mercury Coder is the first commercial diffusion large language model, from Inception Labs (founded by Stanford, UCLA, and Cornell researchers). Instead of emitting tokens one at a time, it refines whole blocks of tokens in parallel across denoising steps, which the company reports makes it 5 to 10 times faster than speed-optimized autoregressive models like Claude 3.5 Haiku and GPT-4o Mini at matching quality, and it was rated the fastest model on Copilot Arena at launch.

Released April 30, 2025 with a 128K-token context window, Mercury Coder is a code-specialized model served through Inception's API and aggregators like OpenRouter. The diffusion approach decouples throughput from output length, which suits agentic coding loops that fire many short calls. Caveats: weights are closed and API-only, vendor speed claims are not independently benchmarked here, and Inception has since shipped newer Mercury 2 and 2.5 generations.