Praxis-1 DROID NVFP4

runwayml/praxis-1-droid (single-arm DROID policy) fine-tuned for 5,000 more steps with NVFP4 quantization-aware training (QAT) on the routed MoE expert layers. Everything else (dense layers, shared experts, routers, action expert) is trained and served in bf16.


Base checkpoint	`runwayml/praxis-1-droid`
QAT	NVFP4 on routed experts: FP4 E2M1 values in blocks of 16, E4M3 block scales, one FP32 scale per expert matrix (`amax / 2688`); weights and activations
Embodiment	single arm, cameras `exterior` + `wrist` (same as the base checkpoint)
Weights	bf16 (QAT keeps bf16 master weights), 54.4 GB, 17 safetensors shards (`model.safetensors.index.json`)
Config	`config.json`, identical to `runwayml/praxis-1-droid`

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-droid-nvfp4")

Load path with runwayml/praxis-1-oss (serve it with configs/serve_droid.yaml). It loads in bf16 like the base checkpoint.

To run the routed experts in NVFP4 (experimental, SM120 GPUs only: RTX 5090, RTX PRO 6000 Blackwell), install the nvfp4 extra (pip install -e '.[nvfp4]') and set model.routed_precision: nvfp4 (for example --set model.routed_precision=nvfp4). The routed experts are packed to NVFP4 from these bf16 files at load time, so the full 54.4 GB is downloaded either way; in memory the policy shrinks from 54.4 GB to 17.5 GB. Set model.nvfp4_cache to keep the packed experts (14.4 GB) between loads. See "Optional routed NVFP4 inference" in the code's README.