Today we’re introducing Mercury 3, the most capable diffusion-based LLM released to date. Across a range of benchmarks, Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5, during operating 4–8x faster at 1,000+ tokens per second.
Mercury 3 demonstrates that content diffusion scales. It’s additional than 40x larger than Mercury 1 and is vastly additional intelligent, but fair as fast.
Early-access partners already use Mercury 3 for sound and conversation support, real-time tutoring, data engineering, and web navigation:
Mercury 3 at a glance
Quality: On par or improved than frontier speed-optimized models akin 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 myriad input and $0.70 per myriad output tokens.
At launch, Mercury 3 is 50% off at $0.10 per 1M input and $0.35 per 1M output. Cached tokens are 90% off the input price.
Diffusion at scale
When we launched Mercury 1, it was the archetypal evidence item that diffusion LLMs could be built for business applications. The open inquiry in the AI community was whether diffusion LLMs could measure to vie alongside autoregressive models on complex use cases.
Mercury 3 answers that question: they can. We’ve scaled Mercury additional than 40x in total parameters, and intellect throughout a range of norm benchmarks has risen 3x from Mercury 1 (mini) to Mercury 3. The scaling laws that drove advancement in autoregressive models clasp for diffusion too.
Mercury 3 in production
Early-access partners are already operating Mercury 3 to power latency-sensitive applications specified as browser use agents, interactive education, real-time hunt agents, and client support.
Browser Use Agents
Agents delay on the example at all step, so small delays compound quickly. Mercury keeps all continue fast, whether the delegate is navigating websites, researching throughout sources, or analyzing logs and tracing incidents. Agents complete sooner and have period to try additional approaches before answering.
We ran Mercury 3 inner our manufacturing web-agent harness against frontier models, on 50 live Mind2Web navigation tasks. Mercury 3 completed 50% of them, considerably additional than GPT-6.1-Sol. It did so using the identical token quantity and the identical figure of delegate steps as a frontier model, at a fraction of the per-token cost.
Shuhao Zhang, Cofounder, TinyFish AI
Interactive Education
Learning runs on accelerated feedback loops, and any postpone pulls a pupil out of the material. Mercury reasons through all petition and responds in near genuine time, whether it’s a sound tutor carrying a conversation, an aide grading work, or an app adapting the next inquiry to what the pupil fair got wrong.
Mercury 3 allowed us to provision all pupil instant feedback without giving up reasoning. Hints, marked answers and definitions now arrive rear in concerning a second, alongside norm as fine as or improved than the models we ran before. When a pupil is stuck on a question, that speed keeps them operating alternatively of waiting, and it was a drop-in toggle for our existing stack.
Michael Tong, Co-founder/CTO, General Learning aka RevisionDojo
Real-time hunt agents
Search applications commencement to awareness unresponsive following fair a few seconds delay. Mercury 3 runs accelerated adequate to procedure dozens of documents and motionless answer during the person is engaged.
On astute glasses, an answer that takes five seconds feels broken. In our Sense API tests, Octen hunt took concerning 80ms; example generation was the bottleneck. Mercury 3 delivered the finest balance of speed and norm among the accelerated models we tested, alongside a median generation period of 0.9 seconds and a median end-to-end reply period of 2.2 seconds.
Kuan Zou, Founder & CEO, Octen AI
Customer support
Support agents have to answer quickly, frequently for regulated businesses anywhere a incorrect answer is costly. Mercury keeps responses conversational without sacrificing accuracy.
At Lorikeet, we assistance complex and regulated businesses build a concierge for their customers that plant throughout voice, conversation and email. In conversation and voice, customers need accelerated responses, but we can't colony on accuracy or getting the job done correctly. In first testing, Mercury 3 attack another models on latency during additionally performing higher on evals and in production.
Michael Gribben, Member of Technical Staff, Lorikeet
A real-time merchandise portfolio
Mercury 3 is the flagship example in a widening portfolio of products that allow real-time AI.
Mercury Voice: Mercury Voice is tailored for the extreme low-latency requirements of sound applications and for minimizing period to archetypal answer token (TTFAT), not fair end-to-end latency. It’s already powering agents at Audivi AI, Altur, and OpenCall.
Mercury Decide: Mercury Decide is a decision example built for high-volume classification. It outperforms Jev on benchmarks including JevBench and LegalBench.
Get started
Mercury 3 is accessible today on the Inception API Platform and through OpenRouter, Baseten, and Vercel. It is OpenAI API compatible, so it drops into your existing stack alongside no rewrites. Enterprise deployments assistance dedicated capacity, autoscaling, compliance controls, and configurable data retention.
Try the API alongside 100 myriad liberated tokens
Enterprise evaluations: we'll partner alongside you on workload fit, eval design, and achievement under your serving constraints. Contact us.
