Summary: Overworld published Waypoint-1 on Hugging Face on January 20, 2026 as a real-time interactive video diffusion world model controlled through text, mouse, and keyboard, and released the 2.3B Waypoint-1-Small weights with the WorldEngine stack. It uses 10,000 hours of captioned gameplay plus diffusion forcing and self-forcing with DMD to support interactive generation, reaching about 30 FPS on an RTX 5090. Long-rollout consistency, fidelity, and stability remain below product grade, so it is best read as an open prototype baseline rather than a game-engine replacement.
January 20, 2026 is when the startup Overworld put Waypoint-1 onto the Hugging Face blog: a “real-time-interactive video diffusion model, controllable and prompted via text, mouse, and keyboard.” Hand it some starting frames and what comes back is not a finished clip but a world you steer on the fly — every frame it produces is conditioned on whatever you are feeding in at that moment. The 2.3B-parameter Waypoint-1-Small had its weights published straight to the Hub, and the inference stack, WorldEngine, landed on GitHub at the same time.
Developers have never before been handed a world model in a form they could download and benchmark: this is not a pre-recorded reel but interactive control at 30 FPS on consumer hardware. A day later, the researcher Akshay Gholam Akaliq judged it “a meaningful step toward interactive world models that can actually respond to real-time control.”
The Model, and How You Control It
Waypoint-1 is a video diffusion model that takes its prompts and its steering from text, mouse, and keyboard. Where it parts ways with standard text-to-video models is the control path: those are fine-tuned with brief control inputs, whereas here, in the team’s words, “you can move the camera freely with the mouse, and input any key on the keyboard.” Its API takes a control input per frame — CtrlInput(button=…, mouse=[x, y]) — together with text prompts and optional forced frames. Under the hood sits a frame-causal rectified flow transformer working in a compressed latent space. The pitch is control with zero latency: shift the camera or hit a key, and the very next frame reflects that input instead of sitting out a batched generation pass.
Two Training Stages: 10,000 Hours of Gameplay and Diffusion Forcing
Training happens in two passes. For pretraining, the team used 10,000 hours of gameplay footage that ships with control inputs and captions, applying diffusion forcing: a causal attention mask runs over the sequence while every frame is noised at its own random level and denoised on its own. The trade-off is that training and inference no longer see the same distribution — stretch a rollout long enough and errors pile up while the frames get visibly noisier. Post-training therefore moves to self-forcing paired with DMD (distribution matching distillation) to close that gap, plus one-pass CFG and few-step denoising, which drives the inference cost down to where real time becomes feasible.
WorldEngine: The Engineering That Buys 30 FPS
The piece that turns “real-time” from claim into fact is the WorldEngine inference stack. On an RTX 5090, Waypoint-1-Small (256 tokens per frame) pushes about 30,000 token-passes per second through a single denoising pass — enough for 30 FPS at 4 denoising steps or 60 FPS at 2. The frame-budget math shows why each optimization counts: 30 FPS leaves you roughly 33 milliseconds per frame, and the whole loop of conditioning, attention, and decoding has to land inside that window on every frame. The tricks are proudly low-level: AdaLN feature caching, a static rolling KV cache built on Flex Attention, fused QKV matmuls, and torch.compile configured with fullgraph and max-autotune while dynamic mode stays off. Bottom line: the smaller model alone did not buy 30 FPS; each denoising step was squeezed for every millisecond it had.
Weights on the Hub, and an Ecosystem Ready on Day One
For a first release, the ecosystem arrives remarkably complete. Waypoint-1-Small sits on the Hub under overworld/Waypoint-1-Small, and Waypoint-1-Medium carries a “coming soon” tag. WorldEngine’s code is housed in the Wayfarer-Labs/world_engine repository on GitHub, while a playable demo runs on the company’s own site. Launch day doubled as a hackathon with an RTX 5090 as the prize — an obvious bid to draw in the first players and builders.
The Distance to Genuinely Playable
The drawbacks are just as plain. The 2.3B small model is what shipped first, and Medium is still nowhere; long-rollout consistency in a latent-space model is the exact problem the training pipeline was engineered around — handled, not solved. Community members are already driving it with keyboard and mouse in real time, and the direction earns credit, though fidelity and stability remain short of product grade.
For product teams, the sober reading: take this as the first interactive world model baseline you can self-host — a fit for interactive-story prototypes, agent training environments, and gameplay validation — and not a stand-in for a game engine. The API surface (per-frame control inputs, forced frames, text prompts) is shaped exactly for scripting scenarios or feeding an agent’s actions into a simulated world, and self-hosting keeps your data inside your own infrastructure. The categorical signal is what matters most: world models have stepped out of papers and closed demos into an open, measurable, iterable phase.