ZERITH 2.5 · H1
This is the H1 mid-training backbone checkpoint, with a 23-dimensional state/action interface. These weights support offline inference; direct robot deployment requires a verified deployment bundle with matching coordinate, unit and normalization conventions.
See the usage guide and project repository.
FP32 weights are stored in model.safetensors; their SHA-256 is recorded in manifest.json.
Obtain the external encoders separately using the pinned revisions in dependencies.json.
Checkpoint size
The main checkpoint contains 6,774,604,284 neural parameters (6.775B) in FP32. The safetensors file is 27.10 GB (25.24 GiB). Parameter counts exclude normalization buffers and separately downloaded encoders.
License
CC BY-NC 4.0 applies to the rights Zerith Robotics holds in the main checkpoint (model.safetensors) hosted in this repository. The separate Wan VAE, UMT5 and VGGT encoders in dependencies.json are obtained from their upstream providers and governed by their respective terms. See NOTICE for attribution and upstream terms for the adapted base model.
Noncommercial use, modification, fine-tuning and redistribution are permitted under its terms; commercial-use rights are not granted. Project source code is licensed separately under Apache-2.0.
CC BY-NC 4.0 适用于本仓库主权重(model.safetensors)中 Zerith Robotics 有权许可的部分。dependencies.json 中列出的 Wan VAE、UMT5 和 VGGT 编码器由各上游单独提供,适用各自的许可。
允许按条款署名后的非商业使用、修改、微调和再分发; 不授予商业使用权,也不要求公开微调权重或业务源码。第三方材料保留各自条款。
- Downloads last month
- 10