AI · SEMICONDUCTOR · FACT CHECK
SKD WIREMostly True
Korean AI startup Inception has unveiled Mercury 3, which it describes as its most powerful diffusion-based large language model (dLLM), asserting that its performance matches that of leading models from OpenAI and Anthropic. The claim is based on performance charts published on Inception's official blog.
Inception introduced Mercury 3 as a top-performing diffusion language model, a distinct architectural approach from the conventional autoregressive models used by most frontier AI labs. The company's benchmark charts, posted on its official blog, indicate performance on par with offerings from OpenAI and Anthropic.
The comparison figures originate solely from Inception's own published materials, which means the parity claim rests on the company's self-reported benchmarks rather than independent evaluation. Diffusion-based LLMs remain a less common approach in the industry, and how the architecture performs against mainstream models at scale is still being assessed by the broader AI community.
The launch underscores Inception's bid to compete directly with established US frontier labs despite its considerably smaller profile, positioning the diffusion approach as a viable alternative path to high-performance language models.
The company's performance claims are broadly supported by its published benchmark charts, though they remain self-reported figures that have not been independently validated — making the characterization Mostly True.
