# Three New Robot Models. You Can Download One.

> Markdown twin of https://roboterminal.app/posts/open-robot-models-download-test · https://roboterminal.app/posts/open-robot-models-download-test.md

**Published:** 2026-10-06

Of three recent physical-AI model launches, only Unitree ships downloadable weights, license, fine-tune code and robot datasets today — read model launches by what you can actually run.

Unitree's UnifoLM-WLA-1.0, Runway's Praxis-1 and Reka's Rho-1 all got filed as physical AI models released. Only Unitree's ships weights, an Apache 2.0 license, fine-tune code and G1 datasets today, with a default fine-tune sized for a single 24GB GPU. Read robot-model launches by what you can run.

**Companies:** [Unitree](https://roboterminal.app/companies/unitree), [Runway](https://roboterminal.app/companies/runway), [Reka](https://roboterminal.app/companies/reka), [Standard Bots](https://roboterminal.app/companies/standard-bots)

**Tags:** humanoids, physical-ai, open-source, robot-learning

Three robot-model launches landed between late September and October 5. [Unitree](https://roboterminal.app/companies/unitree) posted the UnifoLM-WLA-1.0-Base checkpoint on September 28 (https://github.com/unitreerobotics/unifolm-wla). [Runway](https://roboterminal.app/companies/runway) introduced Praxis-1 (https://runway.com/research/introducing-praxis-1), which The Robot Report covered on October 2. [Reka](https://roboterminal.app/companies/reka) put out Rho-1 on October 5 (https://reka.ai/news/rho-1-collapsing-the-multimodal-stack). The coverage filed all three under the same heading: physical AI model released.

That heading hides the one thing that matters to anyone who has to make a robot do something next month. Of the three, one is a model you can download, license, and fine-tune today. One is a model partners are testing with weights promised later. One is a research preview you reach by emailing an address. The useful unit for reading a robot-model launch is what you can run. Parameter count and demo footage are downstream of that.

## The release ladder

Robot-model launches fall on a ladder with four rungs, and the press release almost never tells you which rung you are standing on.

Rung one: an announcement with a demo. Video, a diagram, a claim about generality.

Rung two: a research preview by request. There is a model somewhere, access is gated, and the evidence shown is whatever the lab picked.

Rung three: partner pre-release with open weights promised. Named outside companies run it on their hardware, the weights are not public yet, and the date is "coming months."

Rung four: weights, a license, fine-tune code, and the training data, all pointed at a named robot you can buy.

The distance between rung one and rung four is the distance between a paper and a tool. Here is where the three sit.

**Unitree is on rung four.** The README's open-source checklist shows every item checked: training code for action experts, fine-tuning code, LoRA fine-tuning code, the three models (UnifoLM-ER-1, UnifoLM-ER-Flow, UnifoLM-WLA-1.0-Base), and three datasets (UniBot-V1 Challenge, UnifoLM-WBT-Dataset, UnifoLM-Dex1-Dataset). Per-task datasets sit on Hugging Face under names like G1_Dex1_HangCup and G1_WBT_Brainco_Supermarket_Shelf_Organizing. The license line reads "Except where otherwise noted, this project is released under the Apache License 2.0," with third-party components keeping their own licenses. The checkpoint is on Hugging Face (https://huggingface.co/unitreerobotics/UnifoLM-WLA-1.0-Base). The fine-tune doc says "The default config targets a single 24GB GPU" (https://github.com/unitreerobotics/unifolm-wla/blob/main/docs/train_action_expert_en.md). That is the detail I would underline. A 24GB card sits in a workstation. The default run freezes the VLM backbone and trains only the action-expert DiT head and the robot-state projector, and that is the setup the 24GB target describes.

The rollout was staged. September 11, weights for UnifoLM-ER-1 and UnifoLM-ER-Flow. September 20, model modules and action-expert training code. September 28, the WLA-1.0-Base checkpoint and fine-tuning code. Chinese and trade press picked it up in early October (https://www.chatpicture.com/en/blog/unitree-open-sources-its-6b-humanoid-foundation-model-one-model-runs-both-hands-and-feet-and-embodied-ai-starts-keeping-score-by-the-job). And this is not Unitree's first time on this rung. The company open-sourced its UnifoLM-WMA-0 world-model-action framework before this one (https://github.com/unitreerobotics/unifolm-world-model-action).

**Runway is on rung three.** Runway calls Praxis-1 "our first open-weight world action model," and the wording is doing precise work. The weights will be open when the model ships. Right now: "We're actively testing Praxis-1 with early partners across a variety of embodiments, and will release it publicly in the coming months." Named partners running it on their own hardware are Noble Machines, [Standard Bots](https://roboterminal.app/companies/standard-bots) and Ultra. Other select partners can get pre-launch access by contacting Runway's robotics team. Runway CTO Kamil Sindi told The Robot Report (https://www.therobotreport.com/runway-introduces-praxis-1-world-action-model-robotics/) that "Open weights give hardware developers flexibility and control they don't have today." True statement. Also a statement in the future tense.

**Reka is on rung two.** Reka describes it as "a research preview of Rho-1, our 19B omni-reasoning model trained from scratch." Access is contact@reka.ai. AICoder reported no open weights and no public API (https://aicoder.com/news/news-20261006-reka-rho-1-omni-reasoning-model). Reka's own framing is honest about this: "a functional proof-of-concept and an architectural direction, not a finished product."

One of the three is a download.

## What each one is betting on

The rung tells you what you can run. The bet is the other half: why each lab thinks its approach gets to a useful robot first. These are three different bets.

**Unitree: real-robot hours on hardware it sells.** UnifoLM-WLA-1.0 is a 6B-parameter model trained on approximately 2,500 hours of real-robot data, including Unitree Open Datasets and BitRobot-HIW-500. A single model coordinates 64 tasks: 10 whole-body manipulation tasks and 54 tabletop tasks. It supports two-finger grippers and multiple five-finger dexterous hands. Real-robot evaluations are shown on the Unitree G1, with evaluation videos on the project page (https://unigen-x.github.io/unifolm-wla.github.io/).

Under the hood: a UnifoLM-ER-Flow backbone plus an MMDiT action expert. UnifoLM-ER-1 is built on Qwen3-VL-4B and trained on more than 5 million samples. Unitree says ER-1 leads open-source models on seven of 16 multimodal perception and understanding benchmarks. That is Unitree's own reporting, and some of the comparison results come from Unitree's own testing. File it as a claim.

The structural bet is plain. Unitree manufactures the G1. The per-task datasets carry G1 names. The evaluation videos are G1 videos. Open weights on top of hardware you can buy from the same company is a different object from open weights in the abstract. The model becomes a reason to own the robot, and the robot is where Unitree has shown the model working. That is a moat built from data and physical units rather than from a license.

**Runway: web video scale and a sim-to-real correlation number.** Runway's bet comes from the video side. Praxis-1 is built on the same large-scale video pretraining behind Runway's world models. Sindi's framing to The Robot Report: "Most robot policies are bottlenecked by robot data, which is scarce and expensive to collect." Runway's answer is third-person video. The company's position is that policy performance improves as third-person video scales, and it published a figure contrasting web video with teleoperated robot video. Company claim, company chart.

The sharper number is the sim-to-real one. Runway says simulating robot policies inside its world model predicts real-world results with 0.95 correlation. That is also a company claim. If it holds across partner hardware it would be the most useful thing in any of these three launches, because real-robot evaluation is the slow, expensive part of this whole field. A 0.95 figure is the kind of claim that someone outside Runway should test before anyone plans around it.

Sindi also said "a single model can adapt across very different embodiments, from bimanual arms to humanoids, with light fine-tuning." Training covered manipulation tasks from pick-and-place (lifting soda cans) to deformable objects (packing gift bags). Three partners on their own hardware is a reasonable early test of the cross-embodiment claim. It is still three partners.

**Reka: one network for everything, trained on modest compute.** Reka's bet is architectural. Rho-1 understands and generates text, images, and video, reasons, and outputs robot actions, all in one network. Actions are continuous tokens. Future video frames and actions decode from the same latent state. The companion piece is Reka's Inverse Dynamics Model, which infers control signals from raw video so web video can be used as action-labeled pretraining data. Same instinct as Runway on the data side, with a single network as the answer instead of a world model wrapped around a policy.

The compute story is the interesting part. Trained from scratch on 320 H100 GPUs for three months. For a 19B model that spans text, image, video, and action, that is a lean budget, and Reka lists its own limits without being asked: structural drift on long rollouts, native video capped at 672x384, limited temporal grounding, brittle edits.

The robotics evidence is one LIBERO simulation task, one episode emitting seven action channels with a predicted wrist view. AICoder reported that the robotics demos run on LIBERO simulation without success-rate tables. So the robot part of the omni-model is, for now, a single simulated episode. That is consistent with how Reka describes the whole thing. It is an architecture argument with a robotics demo attached, and the argument is the product.

Put the three bets next to each other and the gap in evidence is wider than the gap in ambition. Unitree's evidence is real hardware, a specific robot, task datasets anyone can download, and videos of that robot doing those tasks. Runway's evidence is three named partners and a correlation number that, as published, comes from Runway itself. Reka's evidence is one simulated rollout. Ambition runs the other direction. Reka is trying to collapse the entire multimodal stack into one network. Runway is trying to make web video substitute for teleop. Unitree is trying to make a G1 hang a cup.

The reason the download matters more than the ambition is that a download lets other people generate evidence. Someone with a 24GB card and a G1 can run the default fine-tune this week and post what happened. Nobody outside the partner list can do that with Praxis-1 until the weights ship, and nobody outside a Reka collaboration can do it with Rho-1. Every claim in the Runway and Reka launches stays a company claim until the rung changes. Some of Unitree's claims are company claims too, the benchmark numbers in particular. But the model itself is sitting there, and that converts the question from "do you believe them" to "did it work when you tried."

## Three objections worth taking seriously

"Unitree only open-sources the model because it sells the robot." Correct, and that is the reason to take the release seriously rather than a reason to discount it. A company whose revenue comes from hardware has a direct interest in outside developers making that hardware more useful. The incentive points toward keeping the weights, code and datasets maintained. It also means the model's best results will cluster on Unitree's own robots, so a team running different hardware should expect to do real work on transfer.

"Promised open weights are close enough to open weights." Runway says Praxis-1 will ship with open weights, and there is no reason to doubt the intent. But a promise has no license text to read, no checkpoint to fine-tune, and no date. A team planning a project around Praxis-1 today is planning around a schedule it does not control.

"Parameter count tells you which model is more capable." Rho-1 is 19B and UnifoLM-WLA-1.0 is 6B. None of the three has published the same evaluation on the same robot, so the sizes cannot be lined up against each other in any meaningful way. Size describes the model. It does not describe what a robot did with it.

## Three checks for the next robot-model launch

**One: can I download the weights today, and under what license?** If the answer is a contact address or "coming months," the launch is on rung two or three, whatever the headline says.

**Two: which robot, and how many real-robot hours?** Unitree names both: the G1 and approximately 2,500 hours. A launch that names neither is asking you to trust the demo reel.

**Three: are per-task success rates on a physical robot published?** Videos show that something worked once on camera. Simulation episodes show the interface works. A success rate across repeated real trials is the number that tells you whether a robot will do the task on a Tuesday.

I think the Unitree launch is the one that moves the field, and not because 6B parameters or 2,500 hours is a large number. It moves the field because the whole stack is Apache 2.0 and shaped to fit a workstation, which means the next hundred results on this model will come from people Unitree does not employ. Runway might end up with the more important idea if the 0.95 holds. Reka might end up with the more important architecture if the drift problems get solved. Neither of those can be checked from outside today. Unitree's can.

The next launch that lands under the physical AI heading will come with a parameter count and a demo reel. The line worth finding first is the license. Unitree's is one sentence in a README. The other two have not published one yet.

## Sources

- [Unitree Robotics, UnifoLM-WLA repository](https://github.com/unitreerobotics/unifolm-wla)
- [Unitree, UnifoLM-WLA project page](https://unigen-x.github.io/unifolm-wla.github.io/)
- [Hugging Face, UnifoLM-WLA-1.0-Base](https://huggingface.co/unitreerobotics/UnifoLM-WLA-1.0-Base)
- [Unitree Robotics, action expert fine-tuning guide](https://github.com/unitreerobotics/unifolm-wla/blob/main/docs/train_action_expert_en.md)
- [Runway, introducing Praxis-1](https://runway.com/research/introducing-praxis-1)
- [The Robot Report, Praxis-1](https://www.therobotreport.com/runway-introduces-praxis-1-world-action-model-robotics/)
- [Reka, Rho-1](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack)
- [AICoder, Rho-1](https://aicoder.com/news/news-20261006-reka-rho-1-omni-reasoning-model)
- [ChatPicture, UnifoLM-WLA coverage](https://www.chatpicture.com/en/blog/unitree-open-sources-its-6b-humanoid-foundation-model-one-model-runs-both-hands-and-feet-and-embodied-ai-starts-keeping-score-by-the-job)
- [Unitree Robotics, UnifoLM-WMA-0 repository](https://github.com/unitreerobotics/unifolm-world-model-action)

## Citation

```
"{quote}"
Source: Roboterminal · {url}
```

---

Source: Roboterminal · https://roboterminal.app/posts/open-robot-models-download-test.md