Inception Announces Diffusion-Based LLM «Mercury 3»

AI startup Inception has announced its diffusion-based language model «Mercury 3». The company claims it possesses the highest performance among diffusion-based language models released to date. In terms of speed, it operates 4 to 8 times faster than OpenAI's «GPT-6 Luna» and Anthropic's «Claude Haiku 5.5».

A «diffusion-based language model» refers to an architecture that differs from conventional large language models (LLMs), which generate text one word at a time sequentially from the beginning. Instead, it generates entire sentences at once through a noise-removal (diffusion) process. This mechanism is theoretically expected to significantly reduce waiting time for response generation, and the Mercury series has attracted attention as a representative example of this approach.

Regarding performance, Inception announced that Mercury 3 outperformed GPT-6 Luna and Claude Haiku 5.5 in 8 out of a total of 12 benchmarks (performance evaluation tests). However, all of these evaluation results were conducted by Inception itself, and at this point, independent verification by third-party organizations has not been performed.

In the AI industry, it is not uncommon for companies to base performance announcements on benchmarks they have conducted themselves. On the other hand, the credibility of such figures tends to be widely recognized only after reproduction testing by independent evaluation organizations or researchers is conducted. For Mercury 3 as well, whether external verification can be obtained becomes a crucial determinant for future evaluation of the model.

The significance of this announcement lies in speed performance. LLM response speed is positioned as a factor that creates practical differences in real-time conversational assistants and business use cases handling large volumes of requests. Should the 4 to 8 times speed difference claimed by Inception be confirmed through independent verification, it could become a catalyst for broader adoption of diffusion-model architectures.

At present, Mercury 3's performance is based on self-reporting by Inception. The next point of attention for readers and industry stakeholders should be how independent researchers and evaluation organizations assess this model, and to what extent its actual performance in user environments aligns with the benchmark results.