Mercury 2.5

• - 🆕 Inception Labs' newest diffusion LLM, successor to Mercury 2.
• - ⚡ Refines many tokens in parallel instead of one-by-one decoding.
• - 📏 260K-token context window, up from Mercury 2's 128K.
• - 🧠 Tunable reasoning depth, so latency scales with thinking needed.
• - 🔧 Native tool calling and schema-constrained structured output.
• - 🌐 Aimed at agent loops, search, voice pipelines, context compaction.
• - 🏢 Proprietary and API-served; no published parameter count.

Inception Labs

Inception Labs builds large language models powered by diffusion rather than the autoregressive generation most LLMs rely on. Instead of producing text one token at a time, its diffusion LLMs (dLLMs) generate many tokens in parallel, which makes them several…

Mercury 2.5 is Inception Labs' newest diffusion large language model (dLLM), a family that departs from standard autoregressive decoding: instead of emitting one token after another, it starts from a rough draft and refines many tokens in parallel over a small number of denoising steps. Inception positions it for production pipelines where many model calls stack up per user action and latency compounds.

Compared with its predecessor Mercury 2, the change is both in capacity and in interface. The context window grows to 260K tokens, against 128K for Mercury 2, and the model adds tunable reasoning levels, native tool use and JSON output constrained to a schema — features aimed at agentic use rather than single-shot completions. Inception reports higher throughput and improved general capability over Mercury 2 in its launch materials; those are vendor-reported characterizations, and independent leaderboard coverage remains limited at the time of writing.

The design intent is to keep diffusion's speed advantage while adding a reasoning mode, so teams can dial thinking up for harder steps without abandoning tight latency budgets. Inception's launch materials cite production uses including context compaction, model routing and tool search inside coding agents.

Practically, Mercury 2.5 suits latency-bound rather than intelligence-bound work: rerankers, voice pipelines, agent loops, query rewriting and summarization, where a fast call is repeated many times around a slower reasoning step. It is API-only and closed-weight, with no disclosed parameter count, license or quantization details.