Inception Labs / inception/mercury-2.5-preview

Mercury 2.5 Preview - access through LLMTR

Mercury 2.5 Preview is the release of Inception Labs' diffusion language model (dLLM) that followed Mercury 2. It uses the same parallel-generation-and-refinement approach as Mercury 2 with double the context window: 260,000 tokens. A single response can be at most 65,536 tokens. It supports tool/function calling and schema-enforced JSON output, and accepts text only. One pricing detail is worth knowing before you switch: its per-token rate is markedly lower than Mercury 2's, but by default it produces roughly five times as many hidden reasoning tokens, and those bill as output. In our measurements the real cost of a short request came out almost identical between the two models. Use the `:none` suffix to turn reasoning off entirely or `:minimal` to keep it light; the default level is `medium`. Only those three levels are advertised on this row, because in measurement `low` did not separate from `minimal` and `high` did not separate from `medium`. This is a preview release: the provider may change its behaviour and its pricing.