Mercury 3 Released as Fastest Diffusion LLM to Date
Today Inception launches Mercury 3, a diffusion-based large language model that sets new benchmarks for speed and capability. The system runs at 1,068 tokens per second on NVIDIA GPUs, outpacing GPT-6 Luna and Claude Haiku 5.5 by four to eight times. Across standard benchmarks, Mercury 3 achieves quality on par with or exceeding those frontier speed-optimized models while maintaining substantially higher throughput.
Scaling Diffusion Models
Mercury 3 scales the architecture more than 40 times larger than Mercury 1 while preserving the speed advantages that made the first generation viable. Intelligence measures across benchmark suites have risen threefold from Mercury 1 (mini) to Mercury 3, demonstrating that scaling laws effective for autoregressive models also apply to diffusion approaches. The 1M token context window represents a fourfold increase over Mercury 2.5's 260K, enabling longer-form tasks without context fragmentation.
Pricing Structure
The per-token pricing runs $0.20 per million input tokens and $0.70 per million output tokens. At launch, Mercury 3 operates at 50 percent off those rates: $0.10 per 1M input and $0.35 per 1M output. Cached input tokens receive 90 percent off the base input price, making repeated-query workloads more economical.
Production Use Cases
Early-access partners have deployed Mercury 3 across several latency-sensitive domains.
Browser use agents completed 50 percent of 50 live Mind2Web navigation tasks, outperforming GPT-6.1-Sol on the same token volume and step count while operating at a fraction of per-token cost. The sustained speed matters because agents wait on the model at every step; small delays compound quickly across long browsing sessions.
Interactive education platforms report that fast feedback loops keep students engaged. Hints, marked answers, and definitions return in about a second with quality matching or exceeding prior models. When a student struggles on a question, the quick response prevents abandonment and allows the app to adapt the next item based on the error.
Real-time search agents on devices such as smart glasses need responses within seconds to feel responsive. In Sense API tests, Octen's search component processed requests in about 80 milliseconds; model generation became the bottleneck. Mercury 3 delivered the best speed-quality balance among fast tested models, with a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds.
Customer support operations in regulated businesses require fast responses without sacrificing accuracy. At Lorikeet, Mercury 3 outperformed other models on latency while also scoring higher on evaluations and in production. The model maintains conversational tone while delivering correct answers, which matters for voice, chat, and email channels.
Associated Product Line
Mercury 3 anchors a product family targeting real-time AI workloads. Mercury Voice targets extreme low-latency voice applications and minimizes time to first answer token, not just end-to-end latency. The model already powers agents at Audivi AI, Altur, and OpenCall. Mercury Decide functions as a decision model for high-volume classification, outperforming Jev on benchmarks including JevBench and LegalBench.
Availability
Mercury 3 is available through the Inception API Platform and distributed via OpenRouter, Baseten, and Vercel. The model uses OpenAI API compatibility, so existing stacks integrate without rewrites. Enterprise deployments offer dedicated capacity, autoscaling, compliance controls, and configurable data retention. The API provides 100 million free tokens for initial experimentation.