Inception Labs / inception/mercury-2

Mercury 2 - access through LLMTR

Mercury 2 is Inception Labs' diffusion language model (dLLM). Instead of emitting tokens strictly one after another it produces several in parallel and refines them step by step, which yields markedly lower latency than comparable models and suits it to latency-sensitive work such as realtime chat, voice turns and high-volume automation. It offers a 128,000-token context window and a single response can be at most 50,000 tokens. It supports tool/function calling (sequential; parallel calls are not available yet) and schema-enforced JSON output via `response_format`. Repeated prompt prefixes are served from a prompt cache, and cached input tokens bill at one tenth the normal input rate. It accepts text only — no image or audio input. The model reasons by default and reasoning tokens are billed as output; the four levels are selected with the `:minimal`, `:low`, `:medium` and `:high` suffixes, and the default is `medium`. Use `:minimal` when latency matters most.