Inception Launches Mercury 3, a High-Speed Diffusion LLM
Mercury 3 delivers 1,000 tokens per second with superior performance over frontier autoregressive models.
1Diffusion Scaling Breakthrough
Inception has released Mercury 3, a diffusion-based LLM that challenges the dominance of autoregressive models. By scaling parameters 40x larger than Mercury 1, the model achieves a 3x increase in benchmark intelligence. This release confirms that diffusion architectures can effectively scale to handle complex, commercial-grade tasks while maintaining high operational speeds.
2Unmatched Inference Velocity
Mercury 3 achieves a throughput of 1,068 tokens per second on NVIDIA GPUs, significantly outpacing GPT-6 Luna and Claude Haiku 5.5. This speed is critical for latency-sensitive applications like voice assistants and real-time search. The model maintains this performance while expanding its context window to 1 million tokens, up from 260K.
3Agentic Navigation Performance
In production tests using the Mind2Web harness, Mercury 3 completed 50% of navigation tasks, outperforming GPT-6.1-Sol. It achieved these results using identical token volumes and agent steps as frontier models. By reducing inference latency, the model allows autonomous agents to attempt more strategies within the same timeframe, improving overall task completion rates.
4Real-Time Educational Impact
The model enables near-instant feedback loops for interactive tutoring systems, preventing student disengagement caused by processing delays. Hints and definitions are delivered in approximately one second, maintaining conversational flow. This drop-in replacement capability allows educational platforms to provide sophisticated reasoning without the typical wait times associated with slower, high-parameter autoregressive models.
5Optimized for Wearable Latency
In Sense API tests for smart glasses, Mercury 3 achieved a median generation time of 0.9 seconds and a total response time of 2.2 seconds. This performance is vital for wearable devices where delays exceeding five seconds render interactions unusable. It currently provides the best balance of speed and quality for real-time search.
6Enterprise Support and Reliability
Lorikeet, an early-access partner, utilizes Mercury 3 for regulated business concierge services across voice, chat, and email. The model maintains conversational accuracy while meeting strict latency requirements. By outperforming other models on production evals, it demonstrates that high-speed diffusion can be safely deployed in industries where response accuracy is financially and operationally critical.
7Specialized Model Portfolio
Inception is expanding its ecosystem with Mercury Voice, optimized for minimal time to first answer token, and Mercury Decide, a classification-focused model. Mercury Decide has already surpassed Jev on benchmarks like JevBench and LegalBench. These specialized variants address specific enterprise needs, from high-volume decision-making to extreme low-latency voice interactions for various partners.
8Integration and Pricing Strategy
Mercury 3 is OpenAI API compatible, allowing for seamless integration into existing stacks without code rewrites. Pricing is set at $0.20 per million input tokens, with a 50% launch discount and 90% off for cached tokens. Enterprise users gain access to dedicated capacity, autoscaling, and configurable data retention controls through the Inception API.
Original article: Hacker News — Inception Launches Mercury 3, a High-Speed Diffusion LLM