California-based AI research and product company Inception has unveiled Mercury 3, its latest diffusion language model, claiming it outperforms speed-optimized rivals GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running four to eight times faster.
The new model generates more than 1,000 tokens per second on widely available NVIDIA GPUs and supports a context window of 1 million tokens, allowing it to process substantially larger amounts of information in a single request.
According to Inception, Mercury 3 is more than 40 times larger than its first-generation model, Mercury 1, while maintaining comparable processing speeds.
The company says the results demonstrate that diffusion-based language models can become more capable without sacrificing the speed advantages of their underlying architecture.
Scaling without sacrificing speed
Traditional large language models generate text sequentially, predicting one token at a time. Diffusion language models take a different approach, generating and refining multiple tokens in parallel.
This allows them to produce responses faster, potentially reducing delays in applications that require frequent interactions with a model.
However, developers have traditionally faced a trade-off between model capability and response speed. Smaller models generally respond faster, while more capable systems often require additional processing time.
Inception says Mercury 3 addresses this challenge by delivering performance comparable to or better than recently released OpenAI and Anthropic models across several standard benchmarks while maintaining inference speeds above 1,000 tokens per second.
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec. pic.twitter.com/TYevDl43Lm
— Stefano Ermon (@StefanoErmon) October 19, 2026
The release marks the company’s fourth model since introducing Mercury 1 in June 2025.
Mercury 2 subsequently introduced reasoning capabilities, while Mercury 2.5 added adjustable reasoning and expanded the context window to 260,000 tokens.
Mercury 3 increases that capacity to 1 million tokens, nearly quadrupling the amount of information the model can handle at once.
“Since we launched Mercury 1, the question we’ve heard most is whether diffusion can scale. Mercury 3 is over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second,” said Stefano Ermon, co-founder of Inception.
Ermon added that the company has already begun training its next model.
Faster responses for developers
The company is targeting applications where response delays can affect usability, including interactive tutoring, live translation, writing assistants and automated software agents.
In agent-based systems, models often perform multiple sequential tasks, such as searching websites, analyzing information or examining software logs. Delays at each stage can increase the time required to complete an operation.
Inception says Mercury 3’s faster inference could help reduce these cumulative delays without requiring developers to switch to less capable models.
Early-access customers are already testing the model in production workflows.
Educational technology company RevisionDojo reported that Mercury 3 reduced response times for student feedback, including hints, definitions, and marked answers, to approximately one second.
Mercury 3 is available through the Inception API, Baseten, OpenRouter, and Vercel.
The model is compatible with the OpenAI API format, allowing developers to integrate it into existing applications without extensive code changes.