Reka Merges With Moonvalley to Build Robots That Simulate Before Acting

Reka, the San Francisco AI lab building multimodal models for the physical world, has merged with Moonvalley, the Toronto-based video generation startup known for its copyright-clean Marey model. The deal brings a team of researchers and engineers with pedigrees from Google DeepMind, Meta, Amazon, Microsoft, Wayve, and Runway into Reka's fold, and signals a sharp pivot toward what both companies are calling the "physical AI era."

The Talent Is the Deal

The merger brings Mateusz Malinowski, PhD and Mikołaj Bińkowski, PhD along with a team of former DeepMind, Meta, Amazon, Microsoft, Google, Wayve, and Runway researchers and engineers to Reka. Malinowski, previously a staff research scientist, and Bińkowski, previously a senior research scientist at Google DeepMind, were both key contributors to models that became Google's Veo. That lineage matters: Veo is currently one of the most capable video generation systems in the world, and the people who built its predecessor are now working on something more ambitious.

The acquisition was structured as an all-share exchange deal. Moonvalley was founded in 2023 and had raised $154 million from General Catalyst, Khosla Ventures, and Y Combinator, among others. Reka itself has raised $167 million in total. No cash valuation for the combined entity has been disclosed publicly.

What They Are Actually Building

The centerpiece of the combined effort is a model called WLAM -- short for World Language Action Model. This is a class of AI that goes beyond generating video or answering questions: it tries to understand the physical rules of the world well enough to simulate what will happen next, and then act on that prediction.

The combined team will focus on WLAM, a unique omni model trained on egocentric and other physical world data to perceive and act in the real world by performing realistic simulation for planning. The model can generate high-fidelity multimodal outputs, translate multimodal perception to robotic execution due to its state-of-the-art spatial and temporal awareness, and perform nuanced reasoning over long-form and streaming video.

Think of it as a robot's internal simulator: before a robotic arm reaches for an object, WLAM runs a mental model of the physics involved, checks whether the action is safe, and only then executes. As Malinowski put it, the goal is building models that "simulate motion, physics, and temporal dynamics in ways that enable robots to reason about consequences before they act -- enabling a much safer, more responsible AI."

Why Moonvalley Brought Something Unique to the Table

Moonvalley was not just another video generation startup. Its flagship model, Marey, was built from the ground up on a fully licensed dataset -- a rare distinction in an industry where most competitors train on scraped web content.

Unlike many AI video models that rely on scraped internet content, Marey was trained entirely on licensed material -- including B-roll from independent filmmakers and archives secured through partnerships with platforms like Vimeo. This makes it one of the first commercially safe AI video systems for production environments, reducing the risk of copyright disputes -- a growing concern in Hollywood following lawsuits from major studios.

Moonvalley was founded by former Google DeepMinders and researchers from Meta, Microsoft, and TikTok. Around 80% of Marey's training data came from filmmakers and agencies who intentionally licensed B-roll and footage, rather than having their content scraped unknowingly. That clean data pipeline is now a strategic asset for Reka as it builds physical world models -- the kind of training data that could matter enormously for robotics, where provenance and physical accuracy of footage is critical.

The Strategic Pivot That Led Here

Reka's trajectory over the past two years reads like a startup finding its footing. In 2024, data platform company Snowflake considered acquiring Reka for $1 billion, but the deal fell through. Reka later shifted from developing frontier models to enterprise image and video applications. The company reached a $1 billion valuation as of July 2025 , backed by investors including Nvidia, Snowflake, DST Global, and Cathexis Ventures.

The deal is seen as another example showing that AI model development startups are in a difficult situation. As one report noted, AI model development is expensive and it is becoming increasingly difficult to secure the necessary computing resources. Startups targeting niche markets such as video and image generation also face copyright issues, making it difficult for them to survive independently. For Moonvalley, merging with Reka offered a path to scale its research ambitions without having to raise another large round in an increasingly competitive funding environment.

Four Pillars, One Mission

Post-merger, Reka is organizing itself around four distinct product lines:

• Reka Labs -- core research on world language action models and foundational physical AI
• Infer -- inference infrastructure and API for multimodal AI, built for enterprise reliability
• Reka Vision -- video processing infrastructure to tag, reason over, search, and clip large volumes of video data
• CLARU -- a training data platform for physical AI, covering egocentric video, robotics trajectories, world-model footage, and expert human judgment at scale

The CLARU data platform is particularly interesting. Claru builds the training data behind frontier AI -- egocentric video, robotics trajectories, world-model footage, and expert human judgment at scale. This is the raw material that WLAM needs to learn how the physical world actually behaves, and controlling that pipeline in-house is a significant competitive advantage.

Who Wins, Who Loses

The clearest winners are Reka's research ambitions. Adding a team of ex-DeepMind researchers who built the precursor to Veo -- in an all-stock deal -- is an extraordinary talent acquisition. Dani Yogatama, CEO of Reka, framed it directly: "This team brings world-class expertise in video generation and multimodal models. Together, we're building AI systems that understand and act in the physical world, from simulation to robotics to real-time decision-making."

For Moonvalley's enterprise customers -- Hollywood studios and ad agencies using Marey -- the near-term picture is less clear. The combined company's focus is shifting toward physical AI and robotics, not filmmaking tools. Whether Marey continues to receive dedicated investment or gets folded into a broader research agenda is an open question.

The broader competitive landscape is also affected. Marey's public release had put it in direct rivalry with a growing field of AI video generators -- models like Runway Gen-3, Luma Dream Machine, Pika, and Haiper. With the merger, Moonvalley is effectively exiting that consumer-facing race and betting on a longer-horizon physical AI play instead. That leaves the video generation market slightly less crowded, but removes one of the few copyright-clean alternatives for studios.

What Becomes Possible Now

The combination unlocks something that neither company could credibly pursue alone: a unified system that can both generate physically plausible video and act on that understanding in the real world. Most robotics AI today treats perception and action as separate problems. WLAM is an attempt to collapse that gap.

Reka Labs is developing world language action models and the foundational science behind the next generation of multimodal AI. With Moonvalley's video generation expertise now integrated, the lab has the pieces to train a model that can watch a scene, simulate its physics, and output motor commands -- all in one forward pass. That is the architecture that companies like Physical Intelligence (pi) and Google DeepMind's robotics division are also chasing, which means Reka is now a legitimate player in one of AI's most contested research races.

Reka says more announcements about its work are coming in the next few weeks. Given the talent now assembled under one roof, it would be worth paying attention.