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

Mercury 3, the most advanced diffusion-based language model (LLM) unveiled recently, has shattered benchmarks, outperforming both GPT-6 Luna and Claude Haiku 5.5. Despite being 40 times larger than its predecessor, Mercury 1, and offering 4-8 times faster processing at speeds of 1,000+ tokens per second, Mercury 3 remains remarkably speedy. Early adopters are leveraging Mercury 3 for voice and chat support, real-time tutoring, data engineering, and web navigation applications.

The model provides output quality on par or superior to frontier models, operates at a blistering 1,068 tokens per second using NVIDIA GPUs, and boasts an expanded context of 1 million tokens. Mercury 3 is priced affordably at $0.20 per million input tokens and $0.70 per million output tokens, offering a 50% discount at launch.

Diffusion LLMs were initially perceived as unable to match autoregressive models in complex scenarios, but Mercury 3 has proven these assumptions wrong. With a 3x increase in intelligence across standard benchmarks from Mercury 1 (mini) to Mercury 3, this model demonstrates that diffusion scaling is indeed possible.

Mercury 3's speed, combined with its high-quality output, makes it an ideal fit for latency-sensitive applications such as browser use agents, interactive education, real-time search agents, and customer support. The model can provide instant feedback, keeping students engaged while answering queries in near real time, thereby eliminating delays that often interrupt learning.

In search applications, Mercury 3 can process multiple documents simultaneously without causing delays, even under pressure. Its fast response time makes it suitable for voice, chat, and email support systems, offering conversational yet accurate responses without compromising on quality. Notably, Mercury 3 has outperformed other models in terms of latency and performance in production environments.

Mercury 3 is currently available on the Inception API Platform and through various platforms like OpenRouter, Baseten, and Vercel. It is OpenAI API compatible, allowing seamless integration into existing stacks without any modifications. For enterprise deployments, dedicated capacity, autoscaling, compliance controls, and configurable data retention are available, with the option to collaborate on workload fit, evaluation design, and performance assessment under specific serving constraints.