DreamZero

23B (DROID / AgiBot)

A 23B World Action Model built on Wan2.1 video diffusion that jointly predicts future world frames and robot actions, giving 2x better zero-shot generalization than leading VLAs.

🔬 Research🔬 Research only⚖ Open weightsWorld ModelVision-Language-Action modelRobotics foundation model

Parameters

23B

parameters

Release date

17 February 2026

Access:DownloadDeployment:💻 Local

Overview

DreamZero is a World Action Model (WAM) — a hybrid of world modeling and robot policy learning. Instead of the usual observation-to-action mapping of vision-language-action models, DreamZero jointly predicts future world states and actions, using video as a dense representation of how the environment evolves.

The model is built on the Wan2.1-I2V video diffusion backbone (14B) and totals 23B parameters (BF16). It comprises a VAE for visual encoding, a text encoder for language instructions, and proprioceptive robot state input; its output is an autoregressive, joint prediction of future video frames and robot actions.

DreamZero was trained exclusively on the DROID dataset (~75k language-annotated episodes) without massive pretraining, demonstrating effective learning from heterogeneous robot data without highly repetitive demonstrations. The developers report 2x better zero-shot generalization than leading VLAs and ~7Hz closed-loop control. Weights are released under the non-commercial CC-BY-NC-4.0 license.

Classification

World ModelVision-Language-Action modelRobotics foundation model

Access & deployment

Download

Local

Weights: Open weights

Key parameters

🧩 Parameters: 23B

✓ Fine-tuning

📥 Input: video, image, text, robot state data

Technical specification

Parameters

23B

parameters

License

CC-BY-NC-4.0

Hardware requirements

23B BF16 weights (safetensors) on the Wan2.1-I2V video diffusion backbone; ~7Hz closed-loop control after optimizations.

Features:✓ Fine-tuning

Modalities

⬇ Input

videoimagetextrobot_state_data

⬆ Output

videorobot_actionsmotion_trajectories

Capabilities and applications

Native model capabilities

Vision-language-action grounding

The ability of a VLA model to ground visual perception and a language instruction into a concrete physical robot action. The model understands the scene and intent, then generates an executable action sequence, closing the loop from observation to motion.

Category: robotics

World simulation

Model's ability to generate coherent, interactive simulations of physical environments — maintaining geometry, lighting, and physics during exploration.

Category: multimodal

Action conditioning

Controlling model generation via action signals (camera, robot pose, commands, speech) rather than text prompts alone.

Category: multimodal

Zero-shot learning

The model's ability to perform a new task without dataset-specific training or hyperparameter tuning — prediction is produced in a single pass from context.

Category: other

Real-time inference

The model's ability to generate responses with very low latency (>1000 tokens/sec) on specialized inference hardware (e.g. Cerebras WSE), enabling interactive, turn-by-turn collaboration with a human.

Category: coding