Inception "Mercury 3", 1000 tok/s diffusion LLM

Mercury 3 is a diffusion-based large language model that Inception presents as the most capable of its kind, claiming quality on par with or better than frontier speed-optimized autoregressive models (e.g., GPT-6 Luna, Claude Haiku 5.5) while running much faster. The model is scaled more than 40× from Mercury 1, yields roughly 3× higher benchmark intelligence versus the earliest mini release, and supports a 1M-token context window. Measured throughput is about 1,068 tokens/sec on NVIDIA GPUs, and Inception reports launch pricing of $0.20 per million input and $0.70 per million output tokens (introductory 50% discount available). The core argument is empirical: diffusion architectures scale and can match or exceed autoregressive models in both speed and quality when sufficiently large.

Early-access deployments emphasize latency-sensitive applications: browser navigation agents (completed 50% of 50 live Mind2Web tasks, outperforming GPT-6.1-Sol on completion rate), interactive tutoring with sub-second feedback, real-time document search (median generation 0.9s, end-to-end 2.2s), and regulated customer support. A product suite extends Mercury 3 into Mercury Voice (ultra-low time-to-first-token tuning) and Mercury Decide (high-volume classification). Mercury 3 is available via Inception’s API and partners, is OpenAI API compatible, and supports enterprise features like dedicated capacity, autoscaling, and configurable data retention.