Physical Intelligence sells access to its newest robot models through a hosted service, while its older models sit on GitHub for anyone to download. The Pi API beta, announced on 13 October 2026, lets a developer upload robot data, fine-tune a pi model on infrastructure the company manages, and run it on a robot through a hosted gateway.[1] The open route is the openpi repository, which carries pi0, pi0-FAST and pi0.5.[3] This article lays the two side by side using only the company's own pages and paper, and lists what none of them says. For background on the model family, see vision-language-action models.
What the Pi API beta offers
The announcement describes three steps: upload your own data, "launch fine-tuning on compute and infrastructure we manage", and use a hosted gateway to run inference on the fine-tuned model.[1] It frames the problem it is solving as the "ML infrastructure, compute, and complex data pipelines" that adapting a general robot model otherwise needs.[1] Access is through a waitlist where applicants describe what they want to build.[1]
Two deployments are described in some detail. Actor, in Utah, runs an autonomous wheel loader that scoops material into a dump truck in a commercial deployment with no operator in the machine. Almond, in San Francisco, used its Axol dual-arm robot to deburr spacers for its own assembly, with a model produced after four days of iteration through the API.[1] A longer list of companies appears as logos with few or no details. These are curated examples from the vendor's own post, and the post reports no success rates.[1]
One sentence matters for architecture: most early applications "ran with cloud-hosted inference through the API", and some deployments had network connectivity challenges.[1] The post describes no on-robot inference option.
What the open route offers
The openpi repository provides three base models: pi0, a flow-based vision-language-action model; pi0-FAST, an autoregressive variant built on the FAST action tokenizer; and pi0.5, for which the repository supports only the flow matching head.[3] Base checkpoints were pre-trained on more than 10,000 hours of robot data, and the repository adds fine-tuned "expert" checkpoints for ALOHA, DROID and the LIBERO benchmark.[3]
Hardware needs are stated plainly in the README:
| Mode | Memory | Example GPU |
|---|---|---|
| Inference | more than 8 GB | RTX 4090 |
| Fine-tuning (LoRA) | more than 22.5 GB | RTX 4090 |
| Fine-tuning (full) | more than 70 GB | A100 80 GB or H100 |
The README says it was tested on Ubuntu 22.04 and that other operating systems are not supported.[3] Its authors also call the release "an experiment": pi0 "was developed for our own robots", and it "may or may not work for you".[3] The expert checkpoints, they warn, "might not generalize to your particular setup", though the DROID one generalised broadly in their experience.[3]
The repository's LICENSE file is Apache 2.0. It also ships a file named LICENSE_GEMMA.txt containing the Gemma Terms of Use.[3] The pi0 paper says the model's vision-language backbone is initialised from PaliGemma, an open 3-billion-parameter model, with 300M parameters added for the action expert.[5] The README as we read it does not say which files each licence governs, so check that before a commercial plan relies on the weights.
Side by side
| Question | openpi (open) | Pi API beta (hosted) |
|---|---|---|
| Models | pi0, pi0-FAST, pi0.5 | "Latest" models, unnamed |
| Access | Download from GitHub | Waitlist |
| Fine-tuning | Your GPUs, JAX or partial PyTorch | Pi-managed compute |
| Inference | Your hardware | Hosted gateway |
| Price | Not applicable to code and checkpoints; your hardware cost | Not stated |
| Latency, quotas, data terms | Under your control | Not stated |
| Robot coverage | ALOHA, DROID, UR5 and LIBERO examples | No list published |
Why hosted inference is plausible, and what it risks
This section is our interpretation, built on figures from Pi's paper. Pi's pi0 paper reports inference time on a single RTX 4090 with three camera images: 14 ms for image encoders, 32 ms for the observation forward pass, 27 ms for ten flow-matching steps, 73 ms in total on-board. For the mobile robot, where inference ran off-board over Wi-Fi, 13 ms of network latency raised the total to 86 ms.[5]
The reason this matters is that the model does not output one action at a time. It produces a chunk of 50 future actions, and the authors executed chunks open-loop: for 20 Hz robots they ran inference every 0.8 seconds after 16 actions, and for 50 Hz robots every 0.5 seconds after 25 actions.[5] A robot that consumes a chunk for half a second or more has slack to absorb a network round trip of tens of milliseconds. That is our reading of those numbers, not a latency figure for the API, which Pi does not publish.
The risk is the flip side. Open-loop execution means the robot cannot react inside a chunk, and if the connection drops, it has only the actions already received. The post acknowledges connectivity problems at some sites but gives no numbers.[1] For a wheel loader or a food line, how the system fails safe when the network does is a question to put to Pi before signing anything.
Where pi0.7 fits
Pi's pi0.7 page, dated 16 April 2026, describes a steerable model that, per the authors, folded laundry on a bimanual UR5e robot with no laundry data, at a success rate matching the zero-shot rate of expert teleoperators who were new to that robot.[4] Most of its evidence is qualitative video, its charts carry no numeric values in the text, and the page does not say whether weights or code are released.[4] The API post says it offers the "latest" models but does not say that is pi0.7.[1] We therefore do not claim that the API serves it.
What the public pages do not tell you
- Price and limits. No pricing, quotas or usage limits.[1]
- Latency and control rates. No figures for the hosted gateway.[1]
- Success rates. The customer examples carry no measured task success.[1]
- Data terms. Nothing on who owns uploaded data, or whether it trains future models.[1]
- Model version. The newest model is unnamed.[1]
- Independent replication. The pi0.7 results are company-reported with no independent replication mentioned.[4]
How to choose
- Start with openpi if your robot resembles ALOHA or DROID and you have one 24 GB-class GPU. It tells you cheaply whether the approach fits your task.
- Join the waitlist if you need the newest models or have no ML infrastructure, and ask for the missing items above in writing.
- Plan for loss of connectivity either way: decide what the robot does when no new chunk arrives.
- Budget data collection first. Pi's own iteration loop is collect demonstrations, train, test on the robot, then collect more where it failed.[1]
Sources and method
We read the five sources listed on 2026-11-07: the Pi API post, the pi0 open-sourcing post, the openpi README and licence files, the pi0.7 page and the pi0 paper, including its inference-time table. We ran no models and tested no robots. Figures are the authors' reports, and the comparison tables and latency reading are our own.
