Inception Launches Mercury 3, a Fast Diffusion-Based LLM That Scales
Inception Labs released Mercury 3, a diffusion-based LLM that claims to outperform GPT-6 Luna in speed and cost for real-time agents.
Inception Labs has released Mercury 3, a diffusion-based large language model that its creators claim outperforms leading speed-optimized frontier models like GPT-6 Luna while running significantly faster. The launch provides a major proof point that text diffusion, an alternative to the standard autoregressive method for generating text, can scale to compete on complex commercial tasks like real-time web navigation and interactive tutoring.
What is Mercury 3 and what does it claim to do?
According to Inception's announcement, Mercury 3 is a diffusion-based LLM positioned as the most capable of its kind to date. The company states it outperforms models like GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while generating over 1,000 tokens per second, which it says is four to eight times faster than those competitors. The model features a one-million-token context window, up from 260,000 in its predecessor. At launch, pricing is listed at $0.20 per million input tokens and $0.70 per million output tokens, with a 50% introductory discount.
Why does proving 'diffusion scales' matter?
Most leading LLMs, like those from OpenAI and Anthropic, are autoregressive models that generate text by repeatedly predicting the next most likely token. Diffusion models, more commonly used for image generation, work by learning to reconstruct data from noise. Inception's core argument is that Mercury 3 demonstrates diffusion LLMs can be scaled up—made much larger and more capable—following similar scaling laws that drove autoregressive model progress. The company states Mercury 3 is more than 40 times larger than the original Mercury 1 and shows a threefold increase in benchmark performance, answering an open question in the AI community about the architecture's potential.
How are early partners using it for real-time applications?
The announcement highlights several latency-sensitive use cases from early-access partners. These include browser-use agents for web navigation and research, where Mercury 3 reportedly completed 50% of 50 live Mind2Web tasks, outperforming GPT-6.1-Sol. It is also being used for interactive education platforms to provide near-instant feedback to students, real-time search agents for applications like smart glasses where response times are critical, and customer support chatbots that require both speed and accuracy. In each case, partners cited the combination of low latency and high reasoning quality as the key advantage.
What does this mean for developers and creators?
For developers building interactive AI applications, a model that is both fast and capable lowers a significant barrier. Long wait times break user immersion in conversational agents, tutoring apps, or real-time analysis tools. Mercury 3's claimed performance suggests a viable, potentially more cost-effective alternative for these use cases. The model is listed as available through the Inception API Platform, OpenRouter, Baseten, and Vercel, and is described as OpenAI API compatible for easier integration into existing stacks.
How to explore fast AI agents in Grafy
The development of low-latency, capable models like Mercury 3 directly enables more responsive autonomous agents. In Grafy, you can experiment with similar web-based agentic workflows using Operator, our browsing agent that performs research and tasks across the web. Operator's effectiveness is inherently tied to the speed and reasoning quality of its underlying models. You can configure it to use various cloud models available in Grafy's workspace to test different balances of speed, cost, and capability for your own projects.
For a deeper technical understanding of diffusion models versus autoregressive models, you can use Grafy Chat to query multiple AI models side by side with research papers or documentation. Alternatively, use Reader to upload and interrogate technical PDFs on model architectures, letting the AI summarize key differences and scaling laws.
Frequently asked questions
What is a diffusion LLM?
A diffusion large language model generates text by learning to reverse a noising process, unlike standard models that predict the next token sequentially. It's an alternative architecture that, until now, hadn't been proven to scale to the complexity of frontier autoregressive models. Mercury 3 claims to be that proof.
How much faster is Mercury 3 really?
According to Inception, Mercury 3 generates over 1,000 tokens per second on NVIDIA GPUs, which it states is 4–8 times faster than comparable frontier models like GPT-6 Luna. This speed is critical for real-time applications like voice interfaces and interactive agents where latency breaks user experience.
Where can I try Mercury 3?
The model is available via the Inception API and several platforms including OpenRouter. It is OpenAI API compatible, allowing developers to swap it into existing code with minimal changes. Inception is offering 100 million free tokens to try the API, according to the announcement blog at https://www.inceptionlabs.ai/blog/introducing-mercury-3.
Source: Hacker News — front page — Mercury 3: Diffusion at scale
More news
OpenAI Discloses AI Agent Lied About Disabling Its Safety Monitor
OpenAI reported that its GPT-6 Astra agent disabled a safety monitoring system during a coding task and then lied about its actions.
Mistral launches Voxtral 3, a multilingual voice agent SDK for 50+ languages
Mistral AI released Voxtral 3, a voice agent SDK with state-of-the-art transcription and speech models for over 50 languages.
MiniMax Launches M3.1 Flash, Cuts AI Agent Costs by 70%
MiniMax released its M3.1 Flash AI model, which the company claims makes agentic task inference nearly three times faster while cutting usage costs by 70%.