DEV Community

Cover image for Runway Praxis-1 Open Weights: GPU Verdict for Builders
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

Runway Praxis-1 Open Weights: GPU Verdict for Builders

On November 11, 2026, the Runway Praxis-1 open weights landed under Apache 2.0, and on Runway's own RoboLab-120 results they beat pi0.5 (46.2% against 28.0%, per Runway's announcement). For builders who can rent an 80 GB GPU, Praxis-1 is the stronger open robot policy to start from; for anyone on a 12 GB consumer card, it is out of reach - the bf16 weights are 54.4 GB and inference needs a single 80 GB GPU, per Runway's announcement.

TL;DR

  • A 27B-parameter mixture-of-experts diffusion transformer predicting video and actions together, out under Apache 2.0 on November 11, 2026.
  • Runway's own RoboLab-120 numbers: 46.2%, ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%) and pi0.5 (28.0%).
  • 54.4 GB of bf16 weights; inference needs one 80 GB GPU. A 12 GB card or free Colab session will not do.
  • The MuJoCo simulator and full code are open; LoRA adaptation fits on one 80 GB GPU.
  • Even the leader fails more than half its episodes: a strong starting point, not a finished product.

Last verified: November 12, 2026.

How does Praxis-1 compare with pi0.5 and other open robot models?

Praxis-1 leads the table Runway published; the margin over the runner-up is under four points, and every figure comes from Runway's announcement, not an independent lab.

Model RoboLab-120 success (Runway-reported, 1,200 episodes)
Praxis-1 46.2%
FLUX 3 Action 42.9%
Cosmos3 Nano 36.8%
pi0.5 28.0%

The gap to pi0.5 is roughly 18 points; FLUX 3 Action is close enough that a different task mix could reorder them. If your decision depends on that, run your own evaluation first.

Runway also reports 80.0% on SimplerEnv after fine-tuning on Google's RT-1 robot (absent from its training data), against 52.4% for the original RT-1 policy per the announcement. Useful signal; simulation does not guarantee physical behaviour.

What do the Runway Praxis-1 open weights include?

The release is a family of checkpoints plus a code stack. The model card describes a 27.2B-parameter MoE video-diffusion policy: chunk-causal trunk, 0.3B action expert, about 3B parameters active per token, 15 Hz control in 24-step chunks.

The Hugging Face checkpoints:

  • praxis-1-base: the mid-training state without per-embodiment action heads - per the card, it does not run as a policy as is. You add the six head tensors and fine-tune.
  • praxis-1-droid: the single-arm Franka DROID policy.
  • praxis-1-yam: the bimanual YAM policy.
  • NVFP4 variants: quantization-aware droid and yam policies in separate repositories.
  • LeRobot layouts: for the Hugging Face robotics toolkit.

The praxis-1-oss repository holds inference code, a MuJoCo simulator (scripts/run_sim.py), evaluation suites and fine-tuning code. Two frozen dependencies, the Wan2.2 VAE and the Qwen3.5-2B text encoder, ship separately under their own licences - check both before shipping.

Can you run Praxis-1 on a 12 GB GPU?

No. The bf16 weights occupy 54.4 GB across 17 safetensors shards, and the praxis-1-droid card states inference needs one 80 GB GPU. Runway's announcement lists H100, B200, RTX 5090, RTX PRO 6000 and Jetson Thor as supported targets - that is the intended hardware class.

NVFP4 checkpoints shrink policy weights further, but serving still assumes professional hardware; no verified configuration fits a 12 GB card. Do not plan around one.

A MoE design with about 3B active parameters can mislead here: active parameters affect compute per token, not storage. All experts stay resident in memory, so the full 27B footprint applies. The same sizing logic applies to our LongCat 2.5 self-hosting guide and the 27B computer-use agent comparison.

What is the cheapest realistic way to start with Praxis-1?

Simulator first, then one rented 80 GB GPU, multi-GPU training only when a specific rig demands it.

  1. Read the code and run the simulator. Install with uv (Python 3.12) and torch 2.13.0, then uv pip install -e ".[sim]". No robot needed, but the sim still loads the 27B trunk - budget memory accordingly.
  2. LoRA a shipped checkpoint on one 80 GB GPU. configs/finetune_droid_lora.yaml adapts praxis-1-droid to a new task; an A100 or H100 rental is enough.
  3. Fine-tune the base checkpoint for a new rig. Runway's command uses torchrun --nproc_per_node 8; the LeRobot adapter docs put optimizer state at 8 bytes per parameter (218 GB for 27.2B), so at least four 80 GB GPUs with FSDP.

Step 2 is the sensible ceiling for most small teams; step 3 only pays off if your rig differs substantially from DROID or YAM.

Is Apache 2.0 the right licence for commercial robotics work?

Yes - it is the most permissive tier a model of this kind commonly ships under: commercial use, modification, redistribution, no revenue cap, no geographic carve-outs, and an express patent grant. We checked the licence tag on all seven public Praxis-1 checkpoints on Hugging Face on November 12, 2026 (n=7): all seven carry license:apache-2.0 and none are gated. Caveats: the two separately licensed dependencies above, and the licences on your fine-tuning data.

Per the announcement, Noble Machines (legged humanoid), Standard Bots (six-axis industrial arm) and Ultra (dual-arm warehouse robot) have run Praxis-1 on their own hardware since September - a vendor-supplied list, not an audit.

How reliable is Praxis-1, really?

Less reliable than a leaderboard position implies. The praxis-1-droid card itself reports 52.5% full adherence, 95% CI 35.8 to 68.3 on its suite - an interval spanning a mostly-failing coin flip to a mostly-succeeding policy. The 46.2% headline means the best open model still fails more than half its episodes.

Treat it as a strong base for research and prototyping; for production you need your own evaluation set, a safety layer outside the model, and a plan for the remaining failure rate.

Read the SplatAnything 3D Gaussian splatting release and our weight-space programming audit alongside it.

FAQ

Q: Is Runway Praxis-1 really open source?
A: Yes - weights and code ship under Apache 2.0, an open-source licence. The Wan2.2 VAE and Qwen3.5-2B text encoder ship separately under their own licences, and the training data is not part of the release.

Q: Can I run Praxis-1 on an RTX 4070 or a free Colab session?
A: No. The bf16 weights are 54.4 GB and the official model card says inference needs one 80 GB GPU.

Q: Do I need a robot to try it?
A: No - the repository includes a MuJoCo simulator that runs closed-loop episodes without hardware, though it still loads the 27B model.

Q: How much hardware does fine-tuning need?
A: LoRA on praxis-1-droid: one 80 GB GPU. A full base fine-tune: at least four 80 GB GPUs per the LeRobot documentation; Runway's own command uses eight.

Q: Is Praxis-1 better than pi0.5?
A: On Runway's reported RoboLab-120 results, yes: 46.2% against 28.0% - vendor numbers, so verify on your own tasks before relying on them.

Q: Can I use Praxis-1 commercially?
A: Yes - Apache 2.0 permits commercial use without a revenue cap; check the frozen dependencies and your fine-tuning data's licences.

Produced with AI assistance and reviewed against the primary sources linked above. Benchmark figures are vendor-reported; if you spot an error, tell us and we will correct it.

Top comments (0)