kairos

# Kairos

Kairos Platform    ｜    ️ GitHub    |    Hugging Face   |    Model Scope  |    Paper   

----- **Kairos** is a 4B-parameter native cross-embodiment world model for unified understanding, generation, and action prediction. Through a progressive Cross-Embodiment Data Curriculum, it learns from general videos, human behavior, and real-robot interaction to build persistent world representations. Powered by a World-Action Model architecture and hybrid linear temporal memory, Kairos jointly predicts future visual states and executable robot actions, enabling long-horizon reasoning, high-quality generation, and real-time edge-side deployment for embodied AI. ## 0. Latest News Jul 2, 2026: ️ We've open-sourced the Kairos3.1 series, including World-Action Model inference code and three model weights: [Kairos3.1-4B-robot-480P](https://huggingface.co/kairos-agi/Kairos3.1-4B-robot-480P), [kairos-4B-robot-RoboTwin2.0](https://huggingface.co/kairos-agi/kairos-4B-robot-RoboTwin2.0), and [kairos-4B-robot-LIBERO-plus](https://huggingface.co/kairos-agi/kairos-4B-robot-LIBERO-plus). RoboTwin 2.0 and LIBERO-Plus support executable action prediction. Jun 16, 2026: We've released the Kairos technical report on arXiv: [Kairos Technical Report](https://arxiv.org/abs/2606.16533). Feb 26, 2026: We've released Kairos video world model weights, including [kairos-robot-4B-480P](https://huggingface.co/kairos-agi/kairos-sensenova-robot-4B-480P), [kairos-robot-4B-480P-distilled](https://huggingface.co/kairos-agi/kairos-sensenova-robot-4B-480P-distilled), and [kairos-4B-720P](https://huggingface.co/kairos-agi/kairos-sensenova-4B-720P). Dec 18, 2025: Kairos 3.0 was officially introduced, along with the Kairos World Model inference code and [kairos-4B-480P-pretrained](https://huggingface.co/kairos-agi/kairos-sensenova-4B-480P-pretrained). ## 1. Motivation Embodied AI is entering the era of Scaling Laws, yet scaling alone cannot overcome the core barriers to reliable real-world interaction. Heterogeneous embodiment data, weak long-horizon reasoning, separated video-and-action modeling, and edge-side compute constraints still limit the practicality of current world models. Kairos addresses these challenges by unifying cross-embodiment pretraining, persistent world-state modeling, future generation, and executable action prediction within an efficient deployment-aware framework. ## 2. Kairos Framework ### Native Pretraining Paradigm with Cross-Embodiment Data Curriculum Kairos introduces a native world-model pretraining paradigm that learns physical, behavioral, and embodied knowledge from scratch, rather than relying on decoupled post-training adaptation. Through a Cross-Embodiment Data Curriculum (CEDC), the model progressively absorbs knowledge from general videos, human behavior data, and real-robot interaction data. This curriculum enables Kairos to move beyond flat data scaling, building world representations that evolve from passive physical understanding to active task intent and embodied control. ### Unified Understanding–Generation–Prediction Architecture Kairos is designed as a native end-to-end architecture for understanding, generating, and predicting the world within a unified Mixture-of-Transformers framework. Instead of treating world modeling as simple video continuation, Kairos maintains persistent world states, where understanding provides causal interpretation, generation unfolds plausible futures, and prediction produces executable robot actions. This unified design allows the model to connect perception, reasoning, planning, and action in a single intelligence loop. ### ⚡ Hybrid Linear Temporal Memory for Efficient Deployment Kairos introduces a hybrid linear temporal memory mechanism for long-horizon world modeling. By combining Sliding-Window Attention, Dilated Sliding-Window Attention, and Gated Linear Attention, the model captures local dynamics, mid-range interactions, and global causal memory under linear-complexity temporal modeling. Together with deployment-aware system co-design, including efficient kernels, quantization, and token streaming, Kairos supports low-memory, high-throughput inference and lays the foundation for real-time closed-loop deployment. ## ✨ 3. Demos #### Physical–causal consistency Kairos leverages causal CoT and physical laws to transform multimodal inputs into deep task logic. It enables autonomous planning and feasibility analysis, shifting the system from "executing commands" to "understanding intent" for real-world robotic actions. #### Cross-embodiment generalization Unified Cross-Embodiment Generation: A single "brain" that generalizes across single-arm, dual-arm, and dexterous-hand platforms. Kairos enables shared, transferable world knowledge with maximal adaptability. Broad Hardware Support: Native compatibility with Agibot G1, Unitree G1, and Songling PIPER, significantly slashing development costs through zero-shot multi-task generalization. #### Accurate Action Prediction Powered by its World-Action Model, Kairos predicts executable robot action trajectories directly from visual observations and task context. This enables accurate, temporally coherent, and physically grounded action generation for long-horizon embodied manipulation. #### High-efficiency inference Real-time Edge Performance: Industry-leading inference speed with ultra-low resource consumption. Optimized for low-latency, high-reliability deployment across single or multi-GPU embodied systems. ## 4. Model Zoo | Download Links | Model Version | Highlights | |:---:|:---:|:---:| | [HuggingFace](https://huggingface.co/kairos-agi/Kairos3.1-4B-robot-480P) [ModelScope](https://www.modelscope.cn/models/kairos-team/Kairos3.1-4B-robot-480P) | kairos3.1-4B-robot-480P | Embodied foundation model for generation | | [HuggingFace](https://huggingface.co/kairos-agi/kairos-4B-robot-RoboTwin2.0) [ModelScope](https://www.modelscope.cn/models/kairos-team/kairos-4B-robot-RoboTwin2.0)| kairos3.1-4B-robot-RoboTwin2.0 | SOTA performance on 50+ bimanual RoboTwin2.0 tasks| | [HuggingFace](https://huggingface.co/kairos-agi/kairos-4B-robot-LIBERO-plus) [ModelScope](https://www.modelscope.cn/models/kairos-team/kairos-4B-robot-LIBERO-plus)| kairos3.1-4B-robot-LIBERO-plus | SOTA performance on LIBERO-plus | | [HuggingFace](https://huggingface.co/kairos-agi/kairos-sensenova-4B-720P) [ModelScope](https://modelscope.cn/models/kairos-team/kairos-sensenova-4B-720P)| kairos-4B-720P | Supports 720P HD output with enhanced fine-grained detail capture | ## 5. Evaluation ### 5.1 Accuracy Benchmarks ### ⚡ 5.2 Deployment #### 5.2.1 Real-time Inference | GPU | Resolution | Memory(GB) | 1 GPU (s) | 4 GPUs (s) | |:---:|:---:|:---:|:---:|:---:| | NV-A800 | 480P | 23.5 | 11.7 | 3.0 | | NV-RTX5090 | 480P | 13.9 | 11.4 | 5.7 | *(results based on kairos-4B-robot 480p distillation) #### 5.2.2 Benchmark for A800 GPU | Model | Parameters | Memory (GB) | Complexity (PFLOPs) | 1 GPU (s) | 4 GPUs (s) | |:---:|:---:|:---:|:---:|:---:|:---:| | Kairos | 4B | 23.5 | 2.3 | 43.3 | 9.5 | | Cosmos 2.5 | 14B | 70.2 | 156.5 (~70x) | 2526.0 | 687.2 | | Wan 2.2 | 5B | 23.4 | 16.6 (~7x) | 201.0 | 85.0 | | Lingbot | 28B | 46.1 | 347.4 (~160x) | 5525.0 | 1436.0 | *（evaluation setting：TI2V mode with 720P/5s） ## 6. Quick Start ### 6.1 Environment Installation ``` … ``` ### 6.2 Download Kairos Models - Download with huggingface ``` … ``` - Download with modelscope ``` … ``` ### 6.3 Run Inference ``` … ``` ### 6.4 Run Inference in Simulation Environments We provide benchmark-specific instructions and evaluation scripts for the following embodied AI benchmarks. | Benchmark | Description | Guide | | ------------ | ---------------------------------------------------------------------------- | ---------------------------------------------------------- | | RoboTwin 2.0 | Dual-arm manipulation benchmark for evaluating long-horizon embodied control and action prediction. | [benchmarks/robotwin](benchmarks/robotwin/README.md) | | LIBERO-Plus | Long-horizon manipulation benchmark for evaluating generalization across tasks and scenes. | [benchmarks/libero_plus](benchmarks/libero_plus/README.md) | ### 6.5 Kairos Training Kairos training uses Flow Matching with local JSONL video manifests and one target output shape per run. The default recipe samples videos to 832x480, 81 frames, and 16 FPS, resizing and temporally resampling source clips as needed. The default ``trainable_models`` selection is ``dit.video_dit``, and EMA is not applied. Install the optional dependencies and launch the one-GPU default: pip install -r requirements-train.txt bash examples/train_kairos.sh Use NUM_PROCESSES to enable the included DeepSpeed ZeRO-1 launcher: NUM_PROCESSES=8 bash examples/train_kairos.sh The configurable conditions are t2v, ti2v, i2v, and null2v. See [docs/TRAINING.md](docs/TRAINING.md) for the manifest schema, model paths, training arguments, condition probabilities, and checkpoint resume behavior. ## 7. About Us Developed and maintained by the Kairos Team. We specialize in Embodied Intelligence and World Model research, with a mission to build Artificial General Intelligence (AGI) that truly understands the physical world. Our goal is to accelerate the industrialization of embodied technologies and reshape the global landscape of AI competition. ## 8. License Kairos is open-sourced under the Apache License 2.0. Feel free to use, modify, and build commercial products on top of it. Check the [LICENSE](LICENSE) file for the full text. ## 9. Citation If you find our work helpful, please cite us. ``` … ``` ## 10. Acknowledgements We would like to thank the contributors to [Qwen-Image](https://huggingface.co/Qwen/Qwen-Image), [Wan2.1](https://github.com/Wan-Video/Wan2.1), [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio), [FastWAM](https://github.com/yuantianyuan01/FastWAM) and [HuggingFace](https://huggingface.co) for their open-source research contributions. --- ⭐ Star us on GitHub if you find [Kairos](https://github.com/kairos-agi/kairos-sensenova) helpful!