Mercury (Diffusion LLM)

"Mercury: Ultra-Fast Language Models Based on Diffusion." Introduces diffusion-based LLMs (dLLMs) that forecast multiple tokens simultaneously via iterative denoising, rather than sequential autoregressive generation. Mercury Coder Mini achieves 1,109 tokens/sec on H100.

Mercury 2 (February 2026) adds reasoning capability with AA Intelligence Index 25 at ~929 tok/s — roughly 10x faster than comparable autoregressive models at similar quality. A genuinely novel architecture paradigm. By Khanna, Kharbanda, Li, Ermon, Grover, Kuleshov et al.

Provenance: initialization is undisclosed (closed weights, no parameter counts), but the tech report frames Mercury as extending the founders' MDLM from-scratch masked-diffusion pretraining line, "trained on the order of trillions of tokens" of web + proprietary data, and cites no AR-to-diffusion adaptation work — consistent with an in-house pretrain rather than an open-checkpoint conversion.