Imbue has raised $200 million to develop AI models designed to reason and code. The Series B round values the company at over $1 billion and brings its total funding to $220 million.
The AI research lab says it wants to build practical agents that can handle larger goals and work safely in the real world. Its current work includes coding models, research into how large language models learn, and tools for building and debugging AI applications.
From intelligence research to practical models
Imbue, formerly known as Generally Intelligent, launched out of stealth last October. Its initial ambition was to investigate aspects of human intelligence that machines lacked, translate those fundamentals into tasks, and test whether different models could learn to solve them in complex 3D worlds built by the company.
The company’s focus has since shifted toward models it considers useful internally. Coding is one starting point, with the article comparing the intended use to GitHub Copilot and Amazon CodeWhisperer. Imbue argues that its distinction is an emphasis on models that can “robustly reason.”
For Imbue, reasoning means more than producing an answer. The company says an effective agent must handle uncertainty, change its approach when needed, ask questions, collect information, consider possible outcomes, and decide what to do in situations that are difficult to predict.
Why Imbue connects coding with reasoning
Imbue sees code as a way for AI systems to take action on a computer and as a means of improving their reasoning. The company argues that an agent can more reliably fulfill some requests by using code to retrieve or work with information than by trying to assemble the same result without it.
The company also says that training on code can help models learn to reason better, while training without code appears to produce weaker reasoning. These are Imbue’s claims about its approach; the article does not report independent results demonstrating them.
The broader idea resembles Adept’s effort to build AI that can automate software processes. The article also notes that Google DeepMind has explored ways to teach AI to control computers, including learning from people completing instruction-following tasks.
Training for stronger conclusions and actions
Imbue says its models are trained on data intended to reinforce good reasoning patterns. It also uses techniques that spend more compute during inference, the stage when a model produces an answer, in pursuit of what the company describes as robust conclusions and actions.
The lab is training models with over 100 billion parameters and optimizing them against its internal reasoning benchmarks. Parameters are parts of a model learned from training data; together, they help define how it performs tasks such as generating text or code.
This training runs on a compute cluster co-designed by Nvidia that contains 10,000 GPUs from Nvidia’s H100 series. Among those participating in the funding round are the Astera Institute, Nvidia, Cruise CEO Kyle Vogt, and Notion co-founder Simon Last.
Building toward custom AI agents
Imbue is also developing AI and machine learning tools, including prototypes for debugging and visual interfaces built on top of AI models. Alongside that work, it is researching how learning happens in large language models.
The company says it does not plan to turn much of its current research into production products immediately. Instead, it views its tools and models as groundwork for more general-purpose AI and a platform that could let people create custom models.
Its stated longer-term goal is to make systems that understand people’s goals, communicate proactively, and work in the background. Imbue says the new funding will speed up development of AI systems that can reason and code. Whether its approach can deliver agents that work reliably across complex real-world tasks remains to be seen.