π0-FAST
FAST tokenizer released and open-sourced mid-January 2025 (X post by @physical_int, 2025-01-16 approx.); π0-FAST weights open-sourced in openpi on 2025-02-04. Not marked first: no explicit 'first' claim verified.
- Input
- text, image
- Output
- action
- License
- apache-2.0 (openpi code); weights built on PaliGemma, Gemma terms apply
- Verified
- 2026-09-29
How to call it
| Provider | Model id | Endpoint / URL | Docs |
|---|---|---|---|
| GitHub (openpi) | gs://openpi-assets/checkpoints/pi0_fast_base | github.com/Physical-Intelligence/openpi | — |
| Hugging Face (LeRobot port) | lerobot/pi0fast-base | huggingface.co/lerobot/pi0fast-base | docs |
Notable capabilities (2)
- FAST action tokenizer (autoregressive VLA): Frequency-space Action Sequence Tokenization (DCT + BPE) compresses action chunks ~10x, letting an autoregressive VLA learn dexterous high-frequency tasks and train up to 5x faster than diffusion/flow π0. source
- DROID generalist checkpoint (found after launch): pi0_fast_droid runs zero-shot on Franka DROID setups for many table-top instructions (openpi). source
Autoregressive sibling of π0 built on the FAST tokenizer.
Sources: HF blog: π0 and π0-FAST, PI on X, openpi.
Other Physical Intelligence models
π0.7 · π0.5 · π0.6 / π*0.6 · π0 (pi-zero)