Inception Labs releases Mercury 3, a diffusion-based AI model

What you could do with it

• A company building smart glasses could use Mercury 3 to power search queries that return answers in about two seconds while the user is still waiting ¹.
• An education platform could give every student instant feedback on exercises—hints, definitions, and marked answers coming back in roughly one second—so learners stay engaged instead of waiting ¹.
• A business deploying web-navigation agents could run browser-based tasks where Mercury 3 completes half of live navigation challenges, matching frontier models using the same number of tokens and steps ¹.

A new kind of language model

Inception Labs has released Mercury 3, which the company calls the most capable diffusion-based large language model (LLM) available today. The model reportedly matches or beats GPT-6 Luna and Claude Haiku 5.5, while running four to eight times faster and generating more than 1,000 tokens per second on NVIDIA GPUs.

An LLM is an AI system designed to understand and generate human language. Most modern LLMs use an autoregressive approach: they predict one word at a time, left to right, like completing a sentence one token at a time. Diffusion works differently. Borrowed from image generation, a diffusion model starts with static and gradually refines it into coherent text, much like how a blurry photo slowly sharpens into focus.

Making diffusion scale up

When Inception Labs launched its first Mercury model, it proved that diffusion could work for real-world use. The big unanswered question was whether the approach could handle the most demanding tasks, or if diffusion models would always lag behind autoregressive ones as they grew larger.

Mercury 3 appears to answer that question. The new model is more than 40 times larger than Mercury 1, and intelligence across standard benchmarks rose three times over the same span. Inception Labs says the scaling laws that drove progress in traditional models also apply to diffusion.

Real-time applications

Early partners are already using Mercury 3 for applications where speed is critical.

Octen AI, which builds search technology for smart glasses, said Mercury 3 delivered the fastest response among the quick models it tested, with a median generation time of about 0.9 seconds and an end-to-end answer time of 2.2 seconds ¹.

General Learning, an education company, said the model lets it give students instant feedback on exercises without sacrificing reasoning quality, bringing hints and definitions back in roughly one second ¹.

In browser-navigation tests on 50 live tasks, Mercury 3 completed 50 percent of them, outperforming GPT-6.1-Sol while using the same number of tokens and agent steps ¹.

Availability

Mercury 3 can be accessed via the Inception API Platform as well as through OpenRouter, Baseten, and Vercel. Because Mercury 3 works with the OpenAI API, software that already uses it can run without needing to change any code ¹. The standard price is $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount of 50 percent. Inception also offers a separate Mercury Voice model tuned for voice applications and a classification model called Mercury Decide.

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

How good is it

According to the company, Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 across standard benchmarks ¹. It also completed 50 percent of live browser-navigation tasks on the Mind2Web test, outperforming GPT-6.1-Sol while using the same number of tokens and agent steps . The model generates text at over 1,000 tokens per second on NVIDIA GPUs, making it four to eight times faster than comparable models . Higher benchmark scores mean the model handles reasoning and language tasks better, and the company says these results are based on its own evaluations .

Key facts

• Mercury 3 is a diffusion-based LLM from Inception Labs
• It reportedly matches GPT-6 Luna and Claude Haiku 5.5 in quality while running 4–8x faster
• It is over 40 times larger than Mercury 1
• Pricing starts at $0.20 per million input tokens with a 50 percent launch discount
• Available via OpenAI-compatible API on Inception, OpenRouter, Baseten, and Vercel

Sources

1. Diffusion at Scale: Introducing Mercury 3 – Inception