ByteDance Seed

Mercury 2 is the first reasoning model fast enough to pick up the phone, a reasoning diffusion language model built for real-time voice agents. Instead of making callers wait through seconds of dead air while an autoregressive model generates thinking tokens one by one, Mercury 2 uses a diffusion large language model architecture to generate tokens in parallel, decoding 1000+ tokens per second on standard NVIDIA GPUs. That speed is fast enough to run a full reasoning pass and start speaking within the latency budget of a natural conversation, reducing the cost of reasoning from seconds of silence to roughly 300 milliseconds. Mercury models work by corrupting clean text into noise, then training a standard Transformer to reverse the process and predict clean text across all positions simultaneously. Because each denoising pass touches many tokens, generation uses the GPU more efficiently than one-token-at-a-time decoding, making custom-silicon-like speed possible on NVIDIA H100s.