Artificial intelligence startup Inception has announced a revival and completed a new $50 million funding round, led by Menlo Ventures, with additional participation from Microsoft's M12 fund, NVIDIA, Databricks, and Snowflake. Notably, the company's founder and co-founder of DeepMind, Mustafa Suleyman, had sold Inception to Microsoft in 2024, and now the company is back to life.

Inception is now betting on a new architecture called diffusion language models (dLLM). This model differs from traditional autoregressive large models (LLMs) and does not generate content word by word, but instead generates text and code through an "iterative optimization" approach—similar to the principle of diffusion models used in image generation.

Its latest model, Mercury, shows outstanding performance: it can generate over 1,000 tokens per second, far surpassing the speed of current mainstream LLMs such as GPT-5, which generate 40–60 tokens per second.

Currently, Mercury is available for use through platforms such as OpenRouter and Poe, with pricing at $0.25 per million input tokens and $1 per million output tokens, offering significant competitive advantages in both speed and cost. Inception's return marks the full release of the potential of diffusion models in the field of text generation.