Mercury 3 Mercury 3 is a diffusion-based large language model from Inception that generates tokens in parallel at over 1,000 tokens per second. It features a 1 million token context window, outperforms GPT-6 Luna and Claude Haiku 5.5 on quality across benchmarks, and is more than 40 times larger than Mercury 1 while remaining similarly fast. The model is available today through the Inception API as well as Baseten and OpenRouter.
Visit productFeatured media
Loading media…
View on XCosigns Received
“I can confirm Mercury-3 works well in real-world agentic operations. Perfect for https://MAXSIM.cloud Fast mode. Congratulations!”
“It's a good model, sir. In my tests it performs really well as an agentic model. Agents at jev speed. I see lots of usages. I need to try the decision API, too. @WolframRvnwlf @MichelIvan92347”
“Very excited to share Mercury 3. It is huge leap it both quality and speed (>1000 tok/s). Happy to have contributed to this. Do check this out!!!”
“wow. i just tested in on some basic legal tasks in german law and so far, its very fast and accurate. well done!”
