Runway Praxis-1: Free 27B Robot Brain That Beats Nvidia's Cosmos

Runway, the company behind some of the most widely used AI video tools, has given away a robot brain. On November 11, 2026, it released Praxis-1, a 27-billion-parameter model that turns video knowledge into robot movement, with open weights under the Apache 2.0 license. Anyone can download it, run it, fine-tune it, and use it commercially, for free. The model is Runway's first step from generating video about the world to controlling machines inside it, and the company says it beats robot models from Nvidia and Physical Intelligence on Nvidia's own leaderboard. The headline numbers are Runway's own, no independent lab has checked them yet, and the model still fails more than half the episodes on that benchmark. Even so, a free, commercially usable robot policy from a video company is a genuine shift in who gets to build robots.

What Praxis-1 actually is

Runway calls Praxis-1 a "world action model." Most robot policies are trained mostly on robot data: recordings of machines being steered by remote control, which are scarce and expensive to collect. Praxis-1 is built the other way around. It starts from Runway's video foundation model, the same large-scale pretraining behind its general world models, and learns how objects behave, how hands move, and what a task looks like partway through, mostly from ordinary video. A small action expert then reads the video model's internal state as it imagines what comes next and works out the motion that produces it.

Each time the robot acts, the model predicts both a short chunk of future video and 1.6 seconds of motion: 24 action rows at 15 Hz. The robot executes the chunk, sends back what its cameras saw, and calls the model again. The architecture has three parts, according to Runway's research post: a video backbone that predicts what the scene will look like next, an action expert that reads the predicted video features, and thin per-robot projections that translate each robot's poses and actions into a shared format, so one model can control different bodies.

The numbers behind that design: 27.2 billion parameters in total, 54.4 GB in bf16, with about 2.8 billion parameters active for any given token. The backbone holds 26.9 billion parameters, and the action expert is a separate 314-million-parameter transformer. Training ran in four stages, from general video pretraining through embodied-video mid-training to joint video-and-action training and finally embodiment-specific fine-tuning. Runway first showed Praxis-1 on September 30, 2026, and promised a public release "in the coming months." The November 11 release keeps that promise in full: weights, inference code, and fine-tuning code, with support for full fine-tuning and the cheaper LoRA method.

The benchmark numbers, and what they do not say

Runway's headline claim is a 46.2% success rate on Nvidia's RoboLab-120 leaderboard, across 1,200 test episodes. The company publishes the comparison table itself:

Read that table with two caveats. First, every figure in it is Runway's own, labeled "data as of November 2026," and no one outside the company has independently checked the results. Second, 46.2% also means the model fails 53.8% of the episodes. That is the state of open robot policies in late 2026: the best of them succeeds less than half the time on a hard benchmark, and the leaderboard is still the honest way to compare them.

There is a third caveat Runway states itself, in its own research post. The "top world action model" claim is scoped to world action models. On the same RoboLab-120 leaderboard, the top two entries are large language model agents, at 86.0% and 57.1% success. Those agents are slower and far more expensive per task, which is exactly the trade-off the next section covers, but the leaderboard crown comes with an asterisk Runway does not hide.

The second benchmark is more striking. Fine-tuned on Google's RT-1 robot, which Runway says never appeared in its training data, Praxis-1 reached 80.0% success on SimplerEnv's official visual-matching grid, against 52.4% for RT-1's original control software. FLUX 3 Action, fine-tuned on the same RT-1 data with Black Forest Labs' published recipe, reached 39.7%. That is the result that matters most for the "adapts to new robots" claim, because RT-1 is an unseen embodiment rather than one the model trained on.

On a real Franka FR3 arm, fine-tuned on the public DROID dataset of about 277 hours of demonstrations, Praxis-1 handles everyday tasks at real speed: placing a mug upside down on a tray, putting an item in a box and closing the lid, clearing bottles into a bin, closing a drawer. Runway says every clip it published shows the policy in control with no teleoperation, played at 1x speed.

What it costs: free weights, expensive hardware

The model itself costs nothing. The weights, the inference code, and the fine-tuning code are all released under Apache 2.0, including for commercial use. That is the whole pricing story for the software, and it is unusually clean. The cost lives one layer down, in the hardware.

Praxis-1 needs one 80 GB GPU to run, according to Runway's research post. The company lists the hardware it runs on out of the box: data center GPUs like the H100 and B200, workstation cards like the RTX 5090 and RTX PRO 6000, and the Jetson Thor computer inside the robot itself. For a buyer, that is real money. An RTX 5090 carries a $1,999 list price (about £1,502, about C$2,848 at November 12, 2026 rates; Nvidia's own UK list is £1,799), and late-2026 street prices have run far higher, around $6,919 in Tom's Hardware's November 3 price tracker. A data center H100 rents for roughly $3.95 an hour, the figure Runway itself uses in its cost comparison, or about £2.97 / C$5.63 an hour.

The hosted route is cheaper than renting your own card. Reactor, the hosted inference service, runs Praxis-1 at $1.80 an hour (about £1.35 / C$2.56), for both the one-arm DROID checkpoint and the two-arm YAM checkpoint, with latency under 260 ms at the 99th percentile. A faster tier, Praxis-1 Fast, runs on two Nvidia B200 GPUs at 62 ms, but it is in early access and Reactor publishes no price for it yet.

Runway also published its own head-to-head cost test against GPT-6 Astra, OpenAI's frontier model driving the same robot through an agent interface. The test, run on November 9, 2026, covered three tasks on the same bimanual YAM station. Praxis-1 finished in 16 seconds on average against Astra's 4 minutes 13 seconds, 15.6 times faster, and cost $0.04 per completed task against $1.87 for Astra, 51 times cheaper. That comparison is Runway's own, on three tasks it chose, and it flatters Praxis-1 by design. The direction of the result is still worth noting: a specialized policy that predicts motion directly will always be cheaper per action than a general model reasoning through an API.

How to run it

Praxis-1 is available in four places. The weights are on Hugging Face, with native support in Hugging Face's LeRobot robotics library. The inference and fine-tuning code is on GitHub. It also ships in vLLM-Omni, the serving engine, and as hosted inference on Reactor. Runway points developers to its developer site and to hands-on tutorials for post-training the model with Nvidia's NeMo Automodel.

Adapting it to a new robot is the part Runway emphasizes most. With LoRA, the cheaper fine-tuning method, you can adapt the model to your own embodiment, cameras, and tasks without retraining the full model. The company demonstrated the range: the same checkpoint adapted to a legged humanoid, a six-axis industrial arm, a dual-arm warehouse robot, a single Franka arm, and the low-cost SO-101 hobbyist arm. For the SO-101, Runway fine-tuned on 3,956 public community datasets from 1,783 contributors, about 1,260 hours of recordings, and held out 102 whole datasets for validation.

One practical detail matters for anyone planning a deployment. The model's guidance settings, the knobs that control how strongly the predicted video follows the instruction, were tuned for the DROID Franka arm, and Runway says other robots need their own tuning. Latency is about 0.55 seconds per policy call on a single H100 in the default chunked mode, roughly a third of the 1.6 seconds of motion each call returns. An opt-in real-time chunking mode cuts that to about 128 ms, and cuts the time the robot spends holding still roughly fivefold, but it is not the default.

Who is already running it

Three launch partners have been running Praxis-1 on their own hardware since the September announcement: Noble Machines on a legged humanoid, Standard Bots on a six-axis industrial arm, and Ultra on a dual-arm warehouse robot. Runway says one checkpoint adapted to all three with light fine-tuning, and showed the same policy working in a studio and a kitchen without retraining.

The release lands in a crowded week for robot brains. Google DeepMind put Gemini Robotics 2.5 into Mercedes-Benz factories a day earlier, and Nvidia's GR00T 2 teaches robots dexterous skills from human video. The difference is the license. DeepMind's and Nvidia's models ship inside their own stacks and partnerships. Praxis-1 ships as open weights that any robot maker can download, inspect, and adapt. Runway frames that as a deliberate bet on openness, pointing to its role as a founding member of the Cosmos Coalition, the group advancing open world models for physical AI alongside Nvidia.

How Praxis-1 compares

Benchmark figures in this table are each vendor's own, and the tests are not the same, so the rows are a map of the field rather than a ranking. The useful contrast is the license column. Praxis-1 is the only entry here that pairs a leaderboard-topping open-weight model with a published hourly hosted price and a documented path to run it on a hobbyist arm that costs a few hundred dollars. That combination, not any single benchmark number, is what makes the release matter.

Common mistakes when reading this release

The first is treating "free" as "cheap." The weights cost nothing, but the model needs an 80 GB GPU, and the workstation card Runway names lists at $1,999 and sells for far more on the street right now. Budget the hardware before the download.

The second is treating 46.2% as a pass mark. It is the best score among world action models on RoboLab-120, and it still means the model fails most episodes. Robot policies are improving fast, but nobody should plan a deployment around a success rate under half without a human in the loop.

The third is ignoring the asterisk on the leaderboard. Frontier language-model agents score higher on RoboLab-120 than any world action model, at 86.0% and 57.1%. They are slower and far more expensive, which is why a specialized policy wins in practice, but the "number one" claim only holds inside the world-action category.

The fourth is assuming the model will work the same on your robot as on Runway's. The guidance settings ship tuned for one Franka arm, and fine-tuning is embodiment-specific by design. The SO-101 results show adaptation is practical, but it is still work, not a download-and-done step.

The fifth is forgetting that a robot policy moves things in the real world. A wrong video prediction in a video generator is an ugly clip. A wrong action chunk in a robot is a dropped object or a bent arm. Runway's research post is unusually candid about where its model's errors come from, most of them from imagining the wrong future rather than reading the wrong motion off the right one, and that is exactly the failure mode to respect in any deployment.

Frequently asked questions

What is Runway Praxis-1?

Praxis-1 is Runway's open-weight "world action model" for robotics, released November 11, 2026. It is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts short chunks of future video and the robot actions to match, built on the video pretraining behind Runway's general world models.

Is Praxis-1 really free?

The weights, inference code, and fine-tuning code are free under the Apache 2.0 license, including for commercial use. Running it is not free: it needs one 80 GB GPU, or $1.80 an hour through Reactor's hosted service, about £1.35 / C$2.56 at November 12, 2026 rates.

How good is Praxis-1 compared with Nvidia's Cosmos?

On Runway's own figures, Praxis-1 reached 46.2% success on Nvidia's RoboLab-120 leaderboard, against 36.8% for Cosmos3 Nano and 28.0% for Physical Intelligence's π0.5. No independent tests exist yet, and frontier language-model agents score higher on the same leaderboard, so treat the ranking as provisional.

What hardware do I need to run Praxis-1?

One 80 GB GPU, according to Runway. That means a data center card like the H100 or B200, a workstation card like the RTX 5090 or RTX PRO 6000, or a Jetson Thor computer on the robot itself. The model is 54.4 GB in bf16.

Can I fine-tune Praxis-1 for my own robot?

Yes. The release includes fine-tuning code with support for full fine-tuning and LoRA, and Runway demonstrated adaptation to a legged humanoid, an industrial arm, a warehouse robot, a Franka arm, and the low-cost SO-101 hobbyist arm, all from one checkpoint.

Praxis-1 is not the best robot model in the world. It is the best open-weight world action model on the benchmark its maker chose, it costs nothing to download, and it runs on hardware a small team can actually buy or rent. That combination is what will decide whether it matters: if robot makers can take a free video-trained brain and teach it their own machines, the bottleneck in robotics moves from who owns the model to who owns the robot. Runway has just removed the first bottleneck for everyone.

Sources

• Runway, "Praxis-1 Is Now Available," November 11, 2026: release announcement, RoboLab-120 table, hardware list, license terms.
• Runway Research, "Praxis-1: From Video Pretraining to Robot Control": architecture, training stages, latency, cost comparison with GPT-6 Astra, SO-101 fine-tuning.
• Praxis-1 on GitHub (runwayml/praxis-1-oss): inference and fine-tuning code.
• Praxis-1 on Hugging Face: weights with native LeRobot support.
• Reactor, "Praxis-1": hosted inference at $1.80 an hour, under 260 ms p99 latency, Praxis-1 Fast early access.
• Mad Robot, "Runway Praxis-1: free open-weight robot AI model released," November 11, 2026: independent summary, partner embodiments, benchmark caveats.
• Daily AI Brief, "Runway Releases Praxis-1 Robot Model With Open Weights," November 11, 2026: launch summary and hardware support.