Introducing Mercury 3 – Inception
Today we’re introducing Mercury 3, essentially the most succesful diffusion-based LLM launched thus far. Across a variety of benchmarks, Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5, whereas operating 4–8x quicker at 1,000+ tokens per second.
Mercury 3 demonstrates that textual content diffusion scales. It’s greater than 40x bigger than Mercury 1 and is vastly extra clever, however simply as quick.
Early-access companions already use Mercury 3 for voice and chat help, real-time tutoring, knowledge engineering, and net navigation:
Mercury 3 at a look
-
Quality: On par or higher than frontier speed-optimized fashions like GPT-6 Luna and Claude Haiku 5.5.
-
Speed: 1,068 tokens per second on NVIDIA GPUs.
-
Context: 1M tokens, up from 260K for Mercury 2.5
-
Price: $0.20 per million enter and $0.70 per million output tokens.
-
At launch, Mercury 3 is 50% off at $0.10 per 1M enter and $0.35 per 1M output. Cached tokens are 90% off the enter worth.
-
Diffusion at scale
When we launched Mercury 1, it was the primary proof level that diffusion LLMs could possibly be constructed for business functions. The open query within the AI group was whether or not diffusion LLMs might scale to compete with autoregressive fashions on complicated use circumstances.
Mercury 3 solutions that query: they will. We’ve scaled Mercury greater than 40x in complete parameters, and intelligence throughout a variety of ordinary benchmarks has risen 3x from Mercury 1 (mini) to Mercury 3. The scaling legal guidelines that drove progress in autoregressive fashions maintain for diffusion too.
Mercury 3 in manufacturing
Early-access companions are already operating Mercury 3 to energy latency-sensitive functions reminiscent of browser use brokers, interactive schooling, real-time search brokers, and buyer help.
Browser Use Agents
Agents wait on the mannequin at each step, so small delays compound rapidly. Mercury retains every go quick, whether or not the agent is navigating web sites, researching throughout sources, or analyzing logs and tracing incidents. Agents end sooner and have time to strive extra approaches earlier than answering.
We ran Mercury 3 inside our manufacturing web-agent harness in opposition to frontier fashions, on 50 stay Mind2Web navigation duties. Mercury 3 accomplished 50% of them, considerably greater than GPT-6.1-Sol. It did so utilizing the identical token quantity and the identical variety of agent steps as a frontier mannequin, at a fraction of the per-token value.
Shuhao Zhang, Cofounder, TinyFish AI
Interactive Education
Learning runs on quick suggestions loops, and any delay pulls a scholar out of the fabric. Mercury causes by every request and responds in close to actual time, whether or not it’s a voice tutor carrying a dialog, an assistant grading work, or an app adapting the subsequent query to what the coed simply received fallacious.
Mercury 3 allowed us to offer each scholar instantaneous suggestions with out giving up reasoning. Hints, marked solutions and definitions now come again in a few second, with high quality pretty much as good as or higher than the fashions we ran earlier than. When a scholar is caught on a query, that velocity retains them working as an alternative of ready, and it was a drop-in swap for our present stack.
Michael Tong, Co-founder/CTO, General Learning aka RevisionDojo
Real-time search brokers
Search functions begin to really feel unresponsive after just some seconds delay. Mercury 3 runs quick sufficient to course of dozens of paperwork and nonetheless reply whereas the consumer is engaged.
On sensible glasses, a solution that takes 5 seconds feels damaged. In our Sense API assessments, Octen search took about 80ms; mannequin technology was the bottleneck. Mercury 3 delivered the perfect steadiness of velocity and high quality among the many quick fashions we examined, with a median technology time of 0.9 seconds and a median end-to-end response time of two.2 seconds.
Kuan Zou, Founder & CEO, Octen AI
Customer help
Support brokers should reply rapidly, usually for regulated companies the place a fallacious reply is dear. Mercury retains responses conversational with out sacrificing accuracy.
At Lorikeet, we assist complicated and controlled companies construct a concierge for his or her prospects that works throughout voice, chat and e-mail. In chat and voice, prospects want quick responses, however we won’t compromise on accuracy or getting the job carried out appropriately. In early testing, Mercury 3 beat different fashions on latency whereas additionally performing larger on evals and in manufacturing.
Michael Gribben, Member of Technical Staff, Lorikeet
An actual-time product portfolio
Mercury 3 is the flagship mannequin in a widening portfolio of merchandise that allow real-time AI.
-
Mercury Voice: Mercury Voice is custom-made for the acute low-latency necessities of voice functions and for minimizing time to first reply token (TTFAT), not simply end-to-end latency. It’s already powering brokers at Audivi AI, Altur, and OpenCall.
-
Mercury Decide: Mercury Decide is a call mannequin constructed for high-volume classification. It outperforms Jev on benchmarks together with JevBench and LegalBench.
Get began
Mercury 3 is on the market in the present day on the Inception API Platform and thru OpenRouter, Baseten, and Vercel. It is OpenAI API suitable, so it drops into your present stack with no rewrites. Enterprise deployments help devoted capability, autoscaling, compliance controls, and configurable knowledge retention.
-
Try the API with 100 million free tokens
-
Enterprise evaluations: we’ll companion with you on workload match, eval design, and efficiency underneath your serving constraints. Contact us.
