72x Faster Inference and 7-Minute Long Video Generation: Daxiao Robot Open-Sources Kairos 3.0-4B

An open-source embodied-native world model

Daxiao Robot has open-sourced Kairos 3.0-4B, the Kairos World Model 3.0 series of embodied-native world models. As the industry's first open-source embodied-native world model integrating multimodal understanding, generation, and prediction, it is built around physical causal consistency, cross-embodiment generalization, ultra-long interaction, real-time cloud-side generation, lightweight efficiency, and on-device embodiment control.

Kairos 3.0-4B is the world's first world model capable of driving embodied robot control on device. It is also the industry's first embodied world model to achieve a 1:1.5 ratio on the THOR edge platform, meaning video generation time is close to the video duration.

Deployed on the Jetson Thor T5000 platform with up to 517 TFlops of computing power, the model can generate robotic-arm motion in 3D simulation, predict and plan trajectories, and use the THOR platform to drive real robot operation, moving robots from performance to practical work.

In authoritative embodied intelligence benchmarks, Kairos 3.0-4B leads across key indicators. On the A800 GPU Benchmark, its inference speed is 72 times faster than Cosmos 2.5, setting a new performance record for embodied world models.

A native world-model architecture for physical understanding

The embodied intelligence industry still faces scarce and fragmented data. Traditional generative models focus on video generation but lack deep cognition of the physical world, leading to bottlenecks in long-horizon interaction, high deployment cost, and weak physical consistency in state prediction.

Kairos 3.0-4B is fundamentally different from generative models that simply retrofit large language or vision models. It is designed from the architecture level for robots operating in the real world, using physical and causal laws as the basis of cognition and building a unified cross-embodiment world-understanding framework.

Around the three capabilities of understanding the world, generating the world, and predicting the world, Kairos 3.0-4B embeds physical laws and causal chains of thought into the model decision process. It fuses real-robot interaction, structured human behavior, and chain-of-thought text data, improving data reuse efficiency and scale-law efficiency for embodied intelligence.

In a complex interaction test, a robot steadily lifts a tray with water, with the water surface showing natural ripples during motion. After placing the tray on a table, the model autonomously plans the placement of milk and apples and arranges them on the tray in order.

Leading physical causal consistency

In difficult physical interaction scenarios such as pouring water and stacking balancing stones, Kairos 3.0-4B demonstrates leading physical causal consistency among mainstream embodied world models by internalizing physical laws and causal reasoning.

When controlling a robot to pour water from a cup into a sink, Kairos 3.0-4B keeps the water flow stable and the total liquid volume consistent with the cup capacity, matching mass conservation and fluid dynamics. In contrast, other models can generate overly fast water flow or liquid volumes that exceed the real cup capacity.

In the balancing-stone scenario, Kairos 3.0-4B accurately reproduces stone rigidity and mechanical balance. Each stacked stone follows gravity and support constraints, while competing models can produce floating stones, loss of rigidity, or disappearing base stones.

Seven-minute coherent long-horizon dynamic interaction

Long-horizon video generation remains a core bottleneck for embodied intelligence. With its integrated understanding-generation-prediction architecture, Kairos 3.0-4B combines agent technology to achieve a breakthrough in long-horizon video generation.

The Kairos agent can hierarchically parse and structure complex user instructions. Relying on the model's fine-grained prediction of temporal-spatial evolution, physical rules, scene dynamics, and interaction logic, it completes continuous world information and iteratively optimizes through self-reflection.

In a home-scene demo, the robot completes a one-shot autonomous workflow: tidying cups and tissues on a table, entering the laundry area, picking up clothes, opening the washing machine, loading clothes, then moving to the kitchen, opening the refrigerator, taking milk, opening a cabinet and drawer, and preparing breakfast with cereal and milk. The whole process remains continuous and physically plausible.

Lightweight architecture with efficient inference

Kairos 3.0-4B leads in inference efficiency, computing consumption, and deployment adaptability. It achieves real-time cloud-side 1:1 inference and can be deployed on the THOR edge platform for efficient on-device inference.

On the A800 GPU benchmark, the self-developed hybrid time-linear attention operator enables an order-of-magnitude improvement in computing efficiency and inference speed. A ten-second generation task takes only 9.5 seconds, about 72 times faster than Cosmos 2.5, 9 times faster than Wan 2.2, and 151 times faster than Lingbot.

With only 4B lightweight parameters and 23.5GB of memory use, Kairos 3.0-4B significantly lowers the deployment threshold while maintaining high performance. It is compatible with GPUs from NVIDIA, MetaX, Hygon, Biren, and others, meeting low-latency, high-reliability, on-device deployment needs.

One brain, many bodies

Kairos 3.0-4B supports strong cross-embodiment generalization, solving the industry pain point of training a separate model for each robot body. The same brain can adapt to multiple bodies and tasks, including single-arm, dual-arm, and dexterous-hand embodiments, without additional training for each task.

Benchmark-leading performance

Kairos 3.0-4B leads in three authoritative embodied intelligence benchmarks: PAI-Bench-robot, WorldModelBench-robot TI2V, and DreamGen Bench. Its results verify the core advantages of physical-level understanding and an efficient architecture.

As a self-developed native embodied world model from China, Kairos 3.0-4B directly addresses bottlenecks in data, computing power, physical reasoning, and deployment. It can act as an efficient data simulator to scale training data at low cost, and it can also drive robot bodies to complete physical tasks, connecting virtual simulation with physical execution.

The technical results have been released at https://github.com/kairos-agi/kairos-sensenova and https://huggingface.co/kairos-agi/kairos-sensenova-common.