Inception Raises 50 Million to Develop Diffusion Language Models with 10X Speed Improvements
Inception announced that it has raised 50 million dollars to further develop its diffusion-based large language models known as dLLMs.
Created on November 6|Last edited on November 6
Comment
Inception announced that it has raised 50 million dollars to further develop its diffusion-based large language models known as dLLMs. The funding round was led by Menlo Ventures with participation from Mayfield, Innovation Endeavors, NVIDIA’s venture arm NVentures, Microsoft’s M12 fund, Snowflake Ventures, and Databricks Investment. The investment signals strong confidence from major technology investors in Inception’s alternative approach to scaling language model performance and efficiency.
The Problem with Current LLMs
Most large language models today rely on autoregression, a process that generates words one at a time in sequence. This method is computationally expensive and slow, resulting in high latency and high cost during inference. These limitations prevent companies from deploying responsive AI systems and force users to experience lag in chat, voice, or interactive settings.
Diffusion as a New Foundation
Inception’s diffusion models apply techniques similar to those used in image and video generation tools like DALL·E, Midjourney, and Sora. Instead of generating tokens sequentially, diffusion models generate outputs in parallel through a denoising process. This parallel approach enables responses that are up to ten times faster while maintaining accuracy and quality. Inception’s first model, called Mercury, already achieves a five to ten times speed improvement compared to optimized models from OpenAI, Anthropic, and Google.
Performance and Efficiency Benefits
The company reports that Mercury’s efficiency reduces the GPU footprint needed for inference. This means enterprises can run larger models with the same hardware or serve more users without scaling up infrastructure. The design makes dLLMs especially suited for latency-sensitive applications such as real-time voice assistants, live code generation, and adaptive user interfaces.
Statements from Investors and Leadership
Tim Tully, Partner at Menlo Ventures, said Inception’s work proves that diffusion models are not just a research concept but a viable foundation for enterprise AI. Inception CEO and co-founder Stefano Ermon added that while model training has become faster, inefficient inference remains the biggest obstacle to AI adoption. According to Ermon, diffusion modeling provides a scalable solution to deliver frontier-level performance with practical deployment costs.
Future Plans
Inception intends to use the new funding to expand its research and engineering teams and accelerate product development. The company’s next phase will focus on advancing diffusion systems that deliver real-time performance across text, voice, and code applications. If successful, this approach could redefine how high-performance language models are built and deployed across industries.
Add a comment