Overworld: Real-Time World Models for the Edge

Our Investment in Overworld

We're thrilled to lead Overworld's $4.5M pre-seed round with participation from Amplify.LA, Garage Capital, Logan Kilpatrick, and angels hailing from Snowflake and Roblox supporting their work on real-time, local-first world models

Why World Models Matter Now

In the past 15 years, we have collectively explored mobile gaming, virtual reality, UGC gaming platforms, virtual world communities. Today, we expect infinite scene and character possibilities, together with realistic physics - all massively customized. One way we think of this new future state would be user-directed worlds, in which anyone can ultimately create a virtual world or even a game through a series of simple prompt commands and context. We think of world models as a system of models, emanating from physic foundation models exploring causality and visual/aural world models which enable pixel, phoneme, sound effect, and music generation and control.

Notably, the virtual world and gaming playspace is one of the most compelling areas for which develop models and get users excited to create and explore. The demand for more adaptive interactive systems is increasing across games, simulations, and creative tools. Users expect environments that respond to their actions, and creators want systems that scale without producing content line by line. Cloud-first AI struggles under these conditions. As interaction becomes continuous, latency becomes noticeable, costs rise with usage, and maintaining coherence over long sessions becomes difficult.

Most interactive software is still built on static systems. Worlds are authored in advance, and logic is fixed. Generative AI is used to create assets or dialogue offline rather than to operate the system itself. This approach breaks down when intelligence needs to run continuously. World models introduce a different abstraction. The model persists over time, maintains state, and updates the environment in real time. The world itself becomes the system.

From Diffusion Research to Real-Time Systems

Diffusion models are typically designed to generate a single output and stop. Overworld restructured diffusion to operate as a persistent system that updates a world incrementally rather than restarting generation at every step. This required treating diffusion as stateful, preserving world state across time and incorporating player input continuously. Instead of producing isolated snapshots, the model updates frame by frame and runs the system itself.

This shift turns diffusion from a generative process into a causal one. The model learns not just to produce plausible frames, but to maintain continuity across them. Objects move because forces are applied. Collisions prevent interpenetration. Motion has inertia. Changes are durable unless acted upon again. Each update must agree with prior state, player input, and learned dynamics, rather than visually correcting errors after the fact.

Running locally on consumer hardware made this practical. Latency dropped to interactive levels, allowing the system to respond multiple times per second. Local execution removed the dependency on cloud inference and enabled continuous interaction without session resets or hidden state loss.

The architecture pairs two components. A diffusion transformer runs locally and handles short-term visuals, motion, and learned physical dynamics. Training emphasizes physical consistency and causal coherence. Overworld has released an early research preview of this system through Waypoint 1, its first public reference implementation. Waypoint 1 demonstrates a persistent, on-device world that updates as the user moves and acts, where interaction produces lasting effects rather than transient reactions.

You can explore Waypoint 1 here.

The Team

Overworld is built by a team of AI researchers and engineers focused on real-time generative world models who worked together in research at Stability AI, the pioneer in diffusion models. The team works in the open, builds distilled diffusion models that run efficiently on consumer hardware, and develops internal evaluation tools such as OWL Eval to measure concise, repeatable model performance.

Louis Castricato, Co-Founder and CEO, leads product and systems design across the company. Prior to Overworld, he worked at Stability AI, where he served as LLM Lead and later Research Director, heading RLHF efforts across large-scale generative models. He studied mathematics at the University of Waterloo and conducted RLHF research at Brown University. His research spans preference learning for storytelling, efficient question answering, and reasoning in language models.

Shahbuland Matiana, Co-Founder and Head of Research, leads diffusion research and infrastructure. He previously worked as a Research Scientist at Stability AI, where he focused on improving diffusion models with human feedback and building multimodal language model architectures. He studied data science at the University of Waterloo and works across diffusion models, reinforcement learning, and scalable training systems. He has built open-source libraries such as tRLX for RLHF on language models and DRLX for diffusion-based learning, along with tooling for large-scale data annotation.

Looking Forward

We believe the next major platform shift in interactive software will be the move from static worlds to real-time world models. As interaction becomes continuous, systems must maintain state, respond immediately, and operate predictably over long sessions. Overworld's focus on real-time, local-first diffusion world models positions the company to become foundational infrastructure for the next generation of games, simulations, and creative tools.

We're excited to partner with the Overworld team as they bring world models from research into production.

Learn more about Overworld.