What's this about?
Inception claims Mercury 2.5 produces 1,107 tokens per second on common NVIDIA GPUs, supports 260K context and costs $0.20/$0.75 per million input/output tokens before a launch discount. It targets repeated low-latency calls in search, voice and coding subagents.
What exactly is a “diffusion language model”?
An autoregressive model writes one token after another. A diffusion model starts from a rough sequence and refines many positions in parallel, closer to editing a whole draft than typing left to right.
What's the community saying?
HN 127 points and 16 comments.
- Developers were excited about using a very fast model as a router, summarizer or judge inside multi-model systems.
- Users noted it is closed-weight and that training on API submissions is opt-out.
- Skeptics said comparisons emphasise speed against older small models while avoiding direct frontier-quality comparisons.