Inception Releases Mercury 3 Language Model, Outpacing GPT-6 Luna and Claude Haiku 5.5

California-based AI research and product company Inception has announced its latest diffusion language model, Mercury 3, claiming it surpasses the speed-optimized competitors GPT-6 Luna and Claude Haiku 5.5 by achieving speeds four to eight times faster. The new model generates over 1,000 tokens per second on widely available NVIDIA GPUs, supporting a context window of 1,000,000 tokens, allowing it to process large amounts of information in a single request.

According to Inception, Mercury 3's scale exceeds that of its first-generation model, Mercury 1, by over 40 times while maintaining a comparable processing speed. The company stated that these results demonstrate that diffusion-based language models can become more powerful without sacrificing their infrastructure speed advantage.

Advantages of Diffusion Language Models

Traditional large language models generate text sequentially, predicting one token at a time. In contrast, diffusion language models take a different approach by generating and refining multiple tokens in parallel. This allows them to produce responses more quickly, potentially reducing latency in applications that require frequent interaction with the model. However, developers have traditionally faced a trade-off between model capability and response speed. Smaller models typically respond faster, while more powerful systems often require additional processing time. Inception claims that Mercury 3 addresses this challenge, delivering performance on par with or better than recently released models from OpenAI and Anthropic in multiple standard benchmark tests, while maintaining an inference speed of over 1,000

tokens per second.

This release marks the company's fourth model since the launch of 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 this capacity to 1,000,000 tokens, nearly quadrupling the amount of information the model can handle at once. Inception co-founder Stefano Ermon stated, "Since we launched Mercury 1, the most common question we've received is whether diffusion technology can scale.

Mercury 3's scale exceeds 40 times, with significantly enhanced capabilities, while still operating at a speed of over 1,000 tokens per second." Ermon added that the company has already begun training the next model.

Applications of Mercury 3

The company targets applications where response latency could affect usability, including interactive tutoring, real-time translation, writing assistance, and automated software agents. In agent-based systems, models often perform multiple sequential tasks, such as searching websites, analyzing information, or checking software logs. Delays at each stage can add to the time required to complete operations. Inception asserts that Mercury 3's faster reasoning can help reduce these cumulative delays without requiring developers to revert to less capable models. Early access customers have already tested the model in production workflows.

EdTech company RevisionDojo reported that Mercury 3 has reduced response times for student feedback, including prompts, definitions, and tagged answers, to about 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.