Inception Releases Mercury 3, a Diffusion Language Model with a 1M-Token Context

Inception Labs Mercury 3 is a diffusion language model that Inception calls the most capable diffusion-based LLM released to date. Its context window is 1M tokens, up from 260K on Mercury 2.5, and it generates 1,068 tokens per second on NVIDIA GPUs.

Per Inception, the model has more than 40 times the parameters of Mercury 1, and Inception says its intelligence across a range of standard benchmarks has risen threefold from Mercury 1 mini. The company says Mercury 3 is on par with or better than speed-optimized frontier models such as GPT-6 Luna and Claude Haiku 5.5 while running 4x to 8x faster, but it does not name the benchmarks. TinyFish AI cofounder Shuhao Zhang says Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in TinyFish's production web-agent harness, significantly more than GPT-6.1-Sol, using the same token volume and number of agent steps as a frontier model. The model runs on the Inception API platform, OpenRouter, Baseten and Vercel, and it follows the OpenAI API format.

Inception's broader claim is that text diffusion scales: the scaling laws behind autoregressive models hold for diffusion too, and Mercury 3 grew without losing speed. Early-access partners run it for latency-sensitive work such as browser agents, tutoring, real-time search and customer support. New API keys come with 100 million free tokens.

Mercury 2 and Mercury 2.5 are closed to new users; existing customers with usage before October 19 keep access until November 15, 2026, and enterprise customers stay supported under their agreements. Mercury 3 heads a three-model lineup that Inception's price list currently shows at 50% off across the board.

Cached input on Mercury 3 costs 90% less than the input price. Mercury Voice, the variant for voice agents, targets a time to first token under 170 ms, and Mercury Decide is a decision model for high-volume classification.