runwayml/praxis-1-base

Praxis-1 Base

The Praxis-1 base model before embodiment-specific post-training: the chunk-causal mid-training checkpoint (P-base r13, step 526,120), trained on a mix of robot and human video across 18 embodiments. It has no per-embodiment action heads, so it does not run as a policy as is. Use it as the starting point for fine-tuning on a new rig.

Model	27.2B-parameter MoE video-diffusion policy: chunk-causal trunk + 0.3B action expert
Frozen dependencies (not in this repo)	Wan2.2 VAE, Qwen/Qwen3.5-2B text encoder
Embodiment heads	none ("embodiment": "base")
Control	15 Hz, 24-step action chunks
Weights	bf16, 54.4 GB, 17 safetensors shards (model.safetensors.index.json)

Before fine-tuning: add the heads of your rig

The checkpoint has the video backbone and the shared action-expert trunk. It leaves out the six tensors that depend on the action and proprio size of a rig:

• action_expert.action_encoder.w1.{weight,bias}
• action_expert.embed_proprio.{weight,bias}
• action_expert.head.{weight,bias}

config.json lists their shapes under heads. Make a checkpoint of your rig with scripts/init_heads.py in runwayml/praxis-1-oss, then fine-tune it as described in praxis/train/README.md, "Fine-tune the base checkpoint" (configs/finetune_base.yaml). The base config has no action_space, normalization or pose_frame; your rig's checkpoint sets them.

Files

• config.json: praxis model config (model_type: praxis, format_version: 2)
• model-000NN-of-00017.safetensors + model.safetensors.index.json: sharded weights
• SHA256SUMS: checksums for every file above

Usage

from huggingface_hub import snapshot_download

path = snapshot_download("runwayml/praxis-1-base")

Then add heads and fine-tune with runwayml/praxis-1-oss.

Related

• runwayml/praxis-1-base: base model, no action heads; the start of a fine-tune for a new rig
• runwayml/praxis-1-droid: single-arm DROID policy
• runwayml/praxis-1-droid-lerobot: praxis-1-droid in the LeRobot policy layout
• runwayml/praxis-1-yam: bimanual YAM policy
• runwayml/praxis-1-eval: evaluation sets
• runwayml/praxis-1-oss: reference code (inference, serving, fine-tuning)

License

Apache License 2.0. The frozen dependencies (Wan2.2 VAE, Qwen/Qwen3.5-2B) are distributed separately under their own licenses.