Diffusion-Model Startup Inception Raises $50M to Challenge the Autoregressive AI Status Quo

TL;DR

• Inception, a new startup developing diffusion-based AI models, has just raised $50 million in seed funding.

Inception raises $50 million to build diffusion models for code and text

With investment pouring into AI at record levels, there’s rarely been a better moment for researchers with ambitious ideas to branch out on their own. For some, building independently may even offer more opportunity than working inside a major lab.

That’s the case for Inception, a new startup developing diffusion-based AI models that has just raised $50 million in seed funding. The round was led by Menlo Ventures, with participation from Mayfield, Innovation Endeavors, Microsoft’s M12, Snowflake Ventures, Databricks Investment, and Nvidia’s NVentures. AI leaders Andrew Ng and Andrej Karpathy also contributed angel investments.

The company is led by Stanford professor Stefano Ermon, a longtime diffusion-model researcher whose work predates their mainstream breakout. Diffusion models — which iteratively refine outputs rather than generating them word-by-word — underpin popular AI image systems like Stable Diffusion, Midjourney, and Sora. Now, Ermon aims to extend those methods to a broader set of challenges through Inception.

Alongside the funding, Inception released a new version of its Mercury model, designed for software development workflows. Mercury is already integrated into tools such as ProxyAI, Buildglare, and Kilo Code. Ermon says Mercury demonstrates one of diffusion’s biggest advantages: efficiency.

“These diffusion-based LLMs are much faster and much more efficient than what everybody else is building today,” he says. “It’s just a completely different approach where there is a lot of innovation that can still be brought to the table.”

Traditional text-based AI relies on autoregressive models, like GPT-5 and Gemini, which produce responses sequentially — predicting each next word based on the one before it. Diffusion models, by contrast, take a holistic view, iteratively shaping an entire output until it matches the target. While autoregression has dominated language systems, recent research suggests diffusion techniques could be stronger for large-scale text processing or when operating under data and compute constraints — advantages that become meaningful when navigating massive codebases.

Diffusion models also offer greater flexibility in how they use hardware. Unlike autoregressive systems, which must compute in strict sequence, diffusion models can execute many operations in parallel, lowering latency for complex tasks.

“We’ve been benchmarked at over 1,000 tokens per second,” Ermon says. “That’s far beyond what’s possible with existing autoregressive technologies, because our system is built to be parallel. It’s built to be really, really fast.”