Do the big hosts serve robot models? A dated read of eight hosts
Checked on October 26 2026, seven general model hosts were read one by one. Not one of them serves a robot policy over its public API. Hugging Face carries the three most downloaded robot checkpoints, and not one of the 500 rows read from its robotics listing carries hosted inference. NVIDIA's build portal does serve robot models, but only its own. Two small specialists serve robot policies today, thirty models between them, and one of the two publishes prices.
Last verified: 2026-10-26. Every host state on this page was read on that date against the host's own pages. All amounts are US dollar amounts. Every rate carries its party, its source page, and the day it was read.
The eight hosts and what each one showed
An absence claim is worth reading only if it names the surface it was read on and the day it was read. Here is the whole set, one row per host. Each row comes from a host's own catalog or model list, not from an aggregator's claim about that host.
| Host | Surface read | What the read showed | Date |
|---|---|---|---|
| Hugging Face | hosted inference providers and the robotics listing | robot weights hosted and heavily downloaded, none of them served through hosted inference, and no robotics task in the served task index | 2026-10-26 |
| Replicate | collections index | 39 categories, none of them robotics; the robotics and vla category URLs return errors while real categories load | 2026-10-26 |
| fal | model catalog | a generative media catalog of image, video, audio and lipsync work, no robot entries | 2026-10-26 |
| DeepInfra | model directory and its served model list | a dozen-plus category set including a World Model category, 175 served models, no robot policy in the served list | 2026-10-26 |
| Together | model catalog and its serverless model list | a language and media catalog, no robot rows in either surface | 2026-10-26 |
| Fireworks | model catalog | a language and vision catalog, no robot rows | 2026-10-26 |
| A multi-provider router | served model list, read directly | 462 served models, none of them robot policies | 2026-10-26 |
| NVIDIA build | robotics explorer | robot models from the NVIDIA namespace only, with no other lab's models on the page | 2026-10-26 |
Every general host keeps a catalog of its own. Not one of those catalogs held a robot policy on October 26 2026. The eighth row is different, and the next section says why.
Hugging Face: the weights are there, the serving is not
Hugging Face comes first in this question. Among the eight hosts here, it is where the robot weights are.
Download counts show how wanted they are. The base SmolVLA checkpoint showed 163,364 downloads on October 26 2026. OpenVLA-7B showed 131,912. GR00T N1.7 3B showed 72,458.
The serving check is one field on the model record. Pull the robotics listing through the public model API and read the hosted inference marker on every row. Every row came back with that field empty. The same marker, checked by name on the five most wanted checkpoints, was empty on all five: SmolVLA, OpenVLA, pi 0.5, Magma, and GR00T N1.7.
Two other surfaces agree. Hugging Face's task index has no robotics row, so there is no robotics category for a hosted provider to sit in. The inference providers page lists 18 providers, and none of them offers a robotics task.
Hugging Face has its own answer, in the blog post on asynchronous robot inference, read in full on October 26 2026. It teaches you to run a policy server, with the model on your own hardware. The post reports about double the task completion speed from running inference asynchronously. That is a serving technique you run, not an endpoint you call.
So the hub publishes the files and leaves the serving to you. That is a sensible business. It is also the answer to the question most people ask.
The NVIDIA portal carries one namespace and nothing else
NVIDIA's build portal is the one host in this survey that serves robot models over an API. Its robotics explorer, read on October 26 2026, lists three entries: cosmos3-nano, cosmos3-nano-reasoner, and isaac-gr00t-synthetic-manipulation.
All three sit in the NVIDIA namespace. No other lab's robot weights appear on that page. If you go looking for SmolVLA, OpenVLA, or pi 0.5 there, you will not find them, because the page is a shelf for NVIDIA's own models.
This one is worth saying plainly, because the opposite belief is common. The GR00T family shows up there because NVIDIA builds GR00T. Other labs' policies are not on that page, and the read on October 26 2026 found no sign of an outside lab's weights.
The hosts and vendors that do serve robot policies
As read on October 26 2026, four parties serve robot models. Two of them are small specialists, one that takes an account and one that takes an invitation. A third runs a first-party waitlist, and a fourth sells a first-party reasoning model with a published card.
| Party | What it serves | How you get in | Published rates |
|---|---|---|---|
| Abstract Inc | six robot models, pi 0.5, MolmoAct2, Picasso, DreamZero Astra, Rho and Isaac 0.5, on one API | an account and their request shape | yes, dated, per million tokens in and per million actions out |
| Robotic Inference | 24 models from 13 families, GR00T N1.7, OpenVLA, SmolVLA, pi 0.5 and X-VLA among them | an invitation, in batches | none published |
| Physical Intelligence | the pi family, from the lab that built it | a waitlist for a beta | none published |
| Gemini Robotics ER 2, an embodied reasoning model | open, a standard key | yes, dated, per million tokens |
Abstract Inc's card, read on October 26 2026, is the one to look at closely. The pi 0.5 row lists 0.15 per million input tokens, 0.015 per million cached input tokens, and 3.70 per million output actions. That card also gives an hour estimate of 0.85 for an hour of a 30 Hz control loop, which they define as 108,000 actions. On that same card, the Isaac 0.5 row's hour estimate of 4.05 runs about five times the pi 0.5 figure. The whole card is on their pricing page. Their homepage names five of the six; the rate card carries the full set.
Google's card prices the reasoning tier. Read on the same day, input runs 1.00 per million tokens and output 5.00 per million tokens through December 31 2026. Both double on January 1 2027. Caching and batch carry the same date. Two facts sit under that card. Gemini Robotics ER 2 reasons about a scene and plans the next step; it does not run the action loop. And Google's own action-loop model has no public rate. The rate card guide walks every line of it.
Physical Intelligence publishes no prices. Its own site, read on October 26 2026, offers a waitlist for a Pi API beta that opened on October 13 2026. You can upload data and fine-tune there. Robotic Inference publishes no prices either. Its pages sit behind a login, and access runs by invitation in batches while it onboards.
Why the search engines send you to rifles and test rigs
On October 26 2026, two of the three lead queries for this topic did not surface a robot model host on the engines this page read.
Search for "fal.ai robotics" and the rifle rows come up in the same result set as the company. The FAL rifle and the media model company share a string. On the engines read that day, one returned the gun rows for it and another returned the company first, which is why the noise and the answer can sit in the same result set. Search for "replicate vla" and the rows belong to VLA-REPLICA, a low-cost, reproducible benchmark for testing robot policies on real hardware that any lab can rebuild from off-the-shelf parts. It has nothing to do with the platform. The Replicate page explains that collision in full.
Autocomplete shows the same gap from the other side. On October 26 2026, "huggingface inference" returned the query itself plus nine platform-level suggestions: inference providers, endpoints, api, playground, pricing, models, api pricing, api key, inference client. Ask the same box for "huggingface inference vla" and it returns nothing. "replicate vla", "fal.ai robotics", and "deepinfra vla" all returned nothing. The box is dense for real model families and silent for robot ones.
Ranking pages finish the story. Read in full, the five closest pages on this question share one gap. Four of them teach you to run a robot policy yourself, with a training walkthrough, a policy server, or a market summary. The fifth is the list of weights. Not one of them reads the catalogs of the hosts you turn to. That gap is why the question has no public answer. It is also what this page is for.
How to re-check every row in five minutes
Every answer on this page has a shelf life. At least one of the parties above changed state inside 2026, so you should date your own read.
Hugging Face: open the robotics listing, pick the checkpoint you want, and read the inference section on its model page. On the day this page was written, the hosted inference marker was empty on every flagship weight checked.
The category hosts: open the categories index and look for robotics. Open the category URL directly and note what comes back. A real category serves as the control. The control matters. Without it you cannot tell a missing category from a broken page.
The money: open the pricing page, read the card, and write the date next to the number. A rate without a date is a number you cannot use.
NVIDIA: open the robotics explorer and read the namespaces on the entries. A lab's name on that page means NVIDIA hosts it. No name means NVIDIA does not.
Frequently asked questions
Is there any big host that serves robot models?
One of the eight, and only its own. NVIDIA's build portal carries robot models from the NVIDIA namespace. The other seven, the general model hosts, each showed no robot policy in their own catalog or served list on October 26 2026. Beyond those eight, two small parties serve robot policies, and a first-party lab runs a waitlist. Every row in the table above names the page it was read from.
Does Hugging Face serve SmolVLA or OpenVLA?
Not through hosted inference on the day this page was written. The weights are on the hub, and the download counts above show how much pull they have. Hosted inference on the model record was empty on every row read. You run those weights yourself, or you buy the behavior from one of the parties above.
What about Replicate, fal and DeepInfra?
All three were checked on their own pages on October 26 2026, and all three came back without a robot policy. Replicate's collections index carries no robotics category, and the category URLs for robotics and vla return errors while real categories load. fal's catalog is generative media work, image and video and audio. DeepInfra's served list has no robot policy in it, though it does publish a World Model category. The Replicate page has the long version with its own receipts.
Is DeepInfra's World Model category a robot category?
No. The category exists and is named for world models, which generate video from a description of an environment. Its model grid did not render on the day this page was written, so this page does not name what sits inside it. A robot policy takes an observation and returns motor commands. The two names sit close together, and the two shelves do not.
Which party should I use?
It depends on which part of the job you are buying. Run the open weights yourself when you hold a GPU and want no per-call bill. Buy from a specialist when you want a policy running today and your team should not own the serving. Use the reasoning tier when the hard part is planning and scene understanding rather than motion. Each card meters a different thing, so this page does not rank them.
Why do the big hosts skip robot models?
The shape does not fit their meters. A robot policy answers on a fixed loop, many times per second, and it has to answer fast. The general hosts sell predictions, images, video, and text, and they bill for the work they can see. A control loop looks like a job that never ends, with a cold start nobody can plan for. Hugging Face took the other road. It publishes the weights, and its blog post teaches you to run the policy server yourself.
Where this leaves you
Robot model hosting is not an empty market. It is a market with a short list of names in it, and the table above is that list, read on October 26 2026. Seven general hosts hold weights or catalogs and no robot policy. One first-party portal serves its own robot models. Two small parties serve robot policies, and one of them publishes a card you can read today.
If you want a policy running this week, the short list above is where to start. Holding a GPU and some patience gets you the weights as free files, and the cost report does the arithmetic for running them yourself. The alternatives guide compares the first-party beta against the hosted parties, and the pricing units guide converts between the ways the two cards meter.
For what this platform publishes and what it costs, the pricing page is the place to look and signup is the way in. The models index shows what the catalog holds today, and that index beats any page's date, including this one.
Last verified: 2026-10-26