Mercury 3

• 🆕 Inception's flagship diffusion LLM, released October 2026.
• ⚡ Inception reports 1,068 tokens per second on NVIDIA GPUs.
• 📏 One-million-token context window, per Inception's model page.
• 🧠 Tunable reasoning effort for latency-versus-depth control.
• 🔧 Tool calling and structured outputs supported.
• 💬 Diffusion decoding refines token blocks in parallel.
• 🌐 Aimed at agents, search and real-time interactive workloads.
• 🎯 Company-reported 50% completion on 50 live Mind2Web tasks.

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 3 is Inception Labs' flagship diffusion large language model, announced on October 19, 2026. Unlike conventional autoregressive systems that emit one token at a time, diffusion LLMs refine blocks of tokens in parallel, which is the source of the family's throughput characteristics: Inception reports 1,068 tokens per second on NVIDIA GPUs for Mercury 3.

Within the Mercury line, each release layered capability on top of that speed. Mercury 2 introduced reasoning to Inception's diffusion models, and Mercury 2.5 added tunable reasoning effort. Inception describes Mercury 3 as a substantial step up in capability over earlier Mercury releases, and its model page lists a one-million-token context window for Mercury 3 — a large expansion over the prior generation — while the model remains above 1,000 tokens per second.

Feature-wise, the model exposes tool calling, structured outputs and tunable reasoning effort, so developers can trade latency against depth of deliberation on a per-request basis. Web search is available through the serving platform, and the model is positioned for agentic pipelines rather than single-turn chat alone.

For agentic evidence, Inception reports running Mercury 3 in its production web-agent harness on 50 live Mind2Web navigation tasks, where the model completed 50% using comparable token volume and agent steps. These figures are vendor-reported; independent evaluations were not published alongside launch. Inception positions the model for workloads such as browser-use agents, real-time search and interactive assistants, where end-to-end latency matters as much as answer quality.