- 代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_inference_and_cookbook_repository; 最近核验：2026-08-11代码：base-model inference and cookbooks released
权重：three base models released
数据：benchmark links; training corpus not released
许可证：Apache-2.0；Repository code; base model and upstream Qwen terms remain.
依赖：transformers 5.2.0、SGLang、Qwen3.5 custom code、safetensors
硬件/传感器：CUDA GPU; 2B is low-entry option、paper VLA uses RTX4090 and three physical platforms
实机证据：Three author platforms, but 1.1 VLA action stack is not released
独立复现：No RynnBrain 1.1 VLA reproduction found.
限制/反面证据：No 1.1 VLA action training/deployment closure；Training corpus not public
当前建议：高潜待持续验证；The base model is usable for perception and grounding experiments, not yet for reproducing the physical VLA table.
重点标签：具身基础模型主流技术2.0、跨本体VLA、可能成为新主流、近期热门
核验来源：来源 1
- 模型/权重 · verified_hold
审核码：verified_hold; 官方关系：paper_claimed_vla_policy_not_released; 最近核验：2026-08-11代码：1.1 VLA action stack not found
权重：RynnBrain-VLA action checkpoints not found
数据：multi-embodiment training data not public
许可证：not_disclosed；No VLA model artifact is available to license.
依赖：RynnBrain base model、unified 81D action space、SONIC for Unitree G1、robot-specific controllers
硬件/传感器：Unitree G1、Astribot S1、Tianji-Wuji、RTX4090
实机证据：Three-platform author results with 20 trials per main task
独立复现：None found.
限制/反面证据：No action-policy weights；No 1.1 VLA training/deployment code；No multi-embodiment training corpus
当前建议：重点更新跟踪；The physical evidence is important, but the released base models do not reproduce the VLA claims.
重点标签：跨本体VLA、实机强证据、动作闭包未开放、近期热门
核验来源：来源 1 · 来源 2
- 权重 · 已核验官方资源
审核码：verified_official; 官方关系：verified_three_public_nongated_base_model_revisions; 最近核验：2026-08-11代码：released separately
权重：2B, 9B, 122B-A10B base models released
数据：training mixture not released
许可证：Apache-2.0；Each inspected model card declares Apache-2.0; Qwen base terms remain relevant.
依赖：transformers 5.2.0 or SGLang、custom model code、Qwen3.5 tokenizer/processor
硬件/传感器：~4.45GB 2B、~18.84GB 9B、~245.15GB 122B-A10B storage
实机证据：These are embodied base weights, not the RynnBrain-VLA action checkpoints used in the physical table
独立复现：No physical VLA reproduction found.
限制/反面证据：Not VLA policy weights；Training corpus not public；122B storage/compute is high
当前建议：当前重点；Three model scales support grounded-perception experiments, while physical action reproduction remains blocked.
重点标签：具身基础模型主流技术2.0、低算力2B入口、跨尺度、近期热门
核验来源：来源 1 · 来源 2 · 来源 3
RynnBrain 1.1将2B/9B/122B-A10B具身视觉语言模型扩展到接触点和3D接地，并以统一动作空间适配Unitree G1、Astribot S1和Tianji-Wuji。三项长时实机任务各20次，generalist平均成功率91.67%；官方Apache-2.0代码与三档基础模型权重可取，但论文中的RynnBrain-VLA训练代码和动作策略权重未随基础模型发布。