Praxis-1 confirmed to run on RTX PRO 6000 Blackwell, DGX Spark, and DGX Station in addition to RTX 5090; LoRA fine-tuning and Hugging Face checkpoint noted
IFM, Runway, and NVIDIA world action models
Latest: IFM and Runway Release Open World Action Models PAN 2 and Praxis-1
Both models are open-weights with permissive licensing, making physical-AI simulation and robot policy testing more accessible. The reproducibility gaps flagged in Praxis-1's evaluation reporting leave some of its performance claims difficult to verify independently.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
IFM (MBZUAI's Institute of Foundation Models) released PAN 2 on November 10 under Apache 2.0, and Runway released Praxis-1 on November 11 under Apache 2.0. Both are world action models aimed at physical AI and robotics, both open-weights.
PAN 2 combines a multimodal LLM for conceptual planning with a 29B-parameter latent diffusion model that renders the resulting simulation. Ablation results show removing the LLM component collapses full-plan completion from 34.0% to 0.0% and step completion from 75.1% to 8.7%. On benchmarks, PAN 2 achieves 34.00% full-plan completion, ahead of MiniMax-H3 at 32.50%, FLUX 3 at 26.77%, and Seedance 2.5 at 24.49%. It scores 49.38 on Physics-IQ Verified in the leaderboard setting, ranking first among open models, though closed model FLUX 3 scores 51.11. In blinded human comparisons across four open models, PAN 2 ranked second on action adherence, physical plausibility, and world consistency. Reward-based learning on top of flow-matching improved causal coherence by 8.5 points and step completion by 2.8 points, but slightly degraded viewpoint invariance by 3.0 points and physical consistency by 2.5 points. IFM found that including gameplay, animation, and simulation footage in early training was detrimental; a 9B test model trained without it won 58% of blinded votes. Training code is included with the release.
Praxis-1 is a 27B-parameter model that jointly predicts video and actions. Germanidis said it leads the RoboLab open-weights leaderboard and outperforms every other world action model on that benchmark. Germanidis also said Praxis-1 is more than 10x faster and 50x cheaper than GPT-6 Astra on the manipulation tasks Runway tested. The Robot Report noted that simulating robot policies inside Runway's world model achieves a 0.95 correlation with real-world outcomes, comparing favorably with 3D reconstruction-based techniques. On November 13, Runway announced Praxis-1 runs on the SO-101 robot platform, trained on thousands of datasets shared by the LeRobot community. The model runs on NVIDIA GeForce RTX 5090, RTX PRO 6000 Blackwell, DGX Spark, and DGX Station using NVFP4 precision, and can be adapted to robots, cameras, and custom tasks with LoRA fine-tuning. The NVFP4 checkpoint is available on Hugging Face.
A technical audit of Runway's Praxis-1 introduction page found that the reported 16 cm final robot-object placement error metric lacks key experimental details: the page does not identify immutable weights or a checkpoint revision, robot embodiment, hardware, task prompts, finetuning recipe, or per-condition raw results. The audit also noted that the 93 evaluation pairs and plus-or-minus 1 SEM are figure-level reporting rather than 93 independent runs per condition, and that SEM is not a 95% confidence interval.
How it developed
Runway extended Praxis-1, its world action model released November 11 under Apache 2.0 for robotics, to run on the SO-101 robot platform on November 13.
The model was trained on thousands of community-shared LeRobot datasets and runs on a single RTX 5090 with NVFP4 precision; Runway made the checkpoint and setup instructions publicly available.
IFM released PAN 2 on November 10 and Runway released Praxis-1 on November 11, both under Apache 2.0, targeting physical AI and robotics applications.
PAN 2 ablations show its LLM planning is load-bearing: removing it drops full-plan completion from 34% to 0%. Runway's Germanidis claimed Praxis-1 leads the RoboLab open-weights leaderboard, though a technical audit found the evaluation page omits key experimental details including robot embodiment and hardware. NVIDIA's Cosmos 3, released November 5, adds a third open world model to the field.
NVIDIA releases Cosmos 3 open models for interactive world simulation on HuggingFace
The Robot Report covers Praxis-1 ahead of public launch, noting 0.95 correlation with real-world outcomes
Sources
4 more sources
- Nvidia GR00T 2 robot foundation model launch
- Hugging Face's Microduck open-source robot
- Skild AI Launches S1 Robot Foundation Model With Single-Video Learning
- Runway's App Studio and Solaris world model
BackgroundFrontier models, open models, and how they are compared
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.