Phi Zero model and project launched

Introduction to Phi Zero

Phi Zero is a newly introduced physical world model built around a concept known as physical language. Humans naturally abstract predictive structure from visual experience and organize it into language for explicit reasoning. Because standard natural language is often too coarse for complex physical transitions, Phi Zero learns a finer-grained physical language directly from unlabeled, in-the-wild videos through self-supervision.

The framework operates on a reason-then-render paradigm. First, it infers future world evolution as a physical-language sequence. Second, it renders the inferred transitions directly into video. This structured approach creates a reusable interface that supports physically realistic generation, action-conditioned simulation, interactive rollouts, and zero-shot transfer.

Architecture and Components

The technical design of Phi Zero is divided into two primary components that handle tokenization and reasoning respectively.

• Physical Language Tokenizer: A transition-level Q-Former extracts ordered features from adjacent latent video states. Finite Scalar Quantization (FSQ) discretizes these features into physical language, while a diffusion-prior decoder reconstructs the video using the sequence alongside a clean first frame. The component utilizes a Wan2.2 VAE encoder and a Wan2.2-5B diffusion decoder, employing 32 queries per adjacent latent-state transition with first-frame conditioning and a pure-noise warm-up.
• Physical Language Reasoner: An autoregressive Vision-Language Model initialized from Qwen3-VL-4B predicts the physical-language sequence based on the initial frame and textual action intent. The vocabulary is extended with 25K atomic symbols, followed by continued pretraining and motion-rich supervised fine-tuning.

Compact Representation and Efficiency

One of the notable achievements of the tokenizer design is its token compression rate. A four-second, 33-frame video is encoded using just 256 discrete physical-language symbols. By comparison, a dense Wan2.2 VAE representation requires 44,800 continuous tokens for the same duration. This results in a representation that uses 175 times fewer tokens while retaining strong reconstruction quality.

Training Data and Curation

The training pipeline utilizes progressive filtering to convert a massive pool of real-world data into refined subsets. Starting from a 50K-hour in-the-wild real-world pool, filtering reduces the volume to 10K hours for tokenizer pretraining. Combining real and simulated videos yields 5 million four-second clips for tokenizer supervised fine-tuning and reasoner pretraining, concluding with 1 million motion-rich, physically informative clips dedicated to final reasoner supervised fine-tuning.

Benchmarks and Capabilities

Phi Zero demonstrates strong performance across multiple physical and intuitive physics benchmarks, leading selected metrics on IQ-Score, Physics Score, Physics Adherence, and IntPhys2 Overall. Beyond video generation, the model's physical language representation enables downstream applications such as cross-embodiment motion transfer—such as translating human body motion into G1 humanoid movements or hand motions into dexterous robotic hands—as well as sim-to-real transfer, interactive world exploration, action-conditioned driving, and robotics simulation.