Inception

AI · LLM Infrastructure

Launch

Inception launches Mercury 3, a diffusion LLM 4-8x faster than rivals

Its parallel generation process breaks the long-standing trade-off between model intelligence and response speed.

• Inception released Mercury 3, its fourth model since Mercury 1 launched in June 2025, now live on the Inception API, OpenRouter, Baseten, and Vercel AI Gateway.
• Inception pioneered commercial diffusion LLMs, which predict many tokens in parallel per step rather than one token at a time like standard autoregressive models.
• Mercury 3 is more than 40x the size of Mercury 1, expands its context window from 260,000 to 1 million tokens, and still runs above 1,000 tokens per second.
• List pricing is $0.20 per million input tokens and $0.70 per million output tokens, cut 50% at launch to $0.10/$0.35, roughly matching or undercutting GPT-6 Luna and Claude Haiku 5.5.
• Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 on GPQA Diamond, IFBench, and SWE-Bench Atlas, though it trails Haiku 5.5 on DeepSWE v1 and MRCR v2; all figures are vendor-reported.
• Inception says it used swarms of Mercury 3 agents for recursive self-improvement, squeezing out an additional 16% speed gain beyond manual engineering upgrades.
• Mercury 3 challenges the industry assumption that frontier intelligence requires either massive compute-hungry models or small, fast but weaker ones, suggesting diffusion architectures could offer both at once.