code_and_weights · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_inference_and_benchmark_code_with_three_public_robot_model_revisions; 最近核验：2026-07-23代码：Fixed commit 661f93337e85e9a30470b109ca645744a1947a65 contains 583 tracked files for video inference and LIBERO-Plus/RoboTwin benchmark evaluation, but no complete post-training or physical-robot adapter.
权重：Three author model revisions are public and non-gated: Kairos3.1-4B-robot-480P (8.09 GB safetensors), RoboTwin2.0 (20.01 GB) and LIBERO-plus (20.07 GB).
数据：No new training corpus or physical-robot dataset is released.
许可证：Apache-2.0_code_and_author_model_cards；Apache-2.0 covers repository code and is declared on the three author model cards. Qwen, Wan, LIBERO and RoboTwin components and datasets retain upstream terms.
依赖：Python、PyTorch/CUDA container selected for A800/A100, RTX 5090 or Metax C500、transformers 5.2.0、flash-linear-attention 0.4.0、Hugging Face Hub 0.36.0、ModelScope 1.26.0、Qwen2.5-VL-7B-Instruct-AWQ、Qwen3-VL-8B-Instruct、Wan2.1-T2V-14B、LIBERO-Plus or RoboTwin for policy evaluation
硬件/传感器：A800/A100, RTX 5090 or Metax C500、Paper reports 23.5 GB A800 memory or 13.9 GB RTX 5090 memory for 480p 5-second inference、Approximately 8.1 GB base robot checkpoint plus approximately 20 GB per benchmark checkpoint
实机证据：未披露
独立复现：Two users ran video inference on RTX 5090/5090-class or RTX 6000 Ada hardware and reported visible motion/physics failures. No independent LIBERO/RoboTwin score or robot deployment was found.
限制/反面证据：No repository tests or formal release tag；No complete training or post-training closure；No physical-robot adapter；G1 deployment is still in progress according to the maintainer；Independent inference reports show deformation, static motion and physics artifacts；No independent action-policy benchmark reproduction
当前建议：近期视频世界模型主流技术2.0候选·社区部分运行；Run a fixed-prompt 480p inference and one benchmark subset from the pinned artifacts before accepting headline policy scores. Do not infer a physical robot deployment path from video generation.
重点标签：世界模型与VLA主流技术2.0、artifact:code+weights、deployment:simulation-first、community:mixed-runtime-evidence
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5 · 来源 6 · 来源 7 · 来源 8 · 来源 9 · 来源 10
