Praxis-1: A Developer's Checklist Before You Run an Open Robot Model on Your Hardware

Shivam Kumar | November 12, 2026

Runway released Praxis-1 on November 11, 2026 — a 27-billion-parameter world action model with open weights under Apache 2.0, commercial use included. It tops NVIDIA's RoboLab-120 leaderboard at 46.2% success (555 of 1,200 episodes), ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and π0.5 at 28.0%.

The model predicts future video and robot actions together: a 26.9B-parameter MoE diffusion-transformer video backbone plus a ~314M-parameter action expert. Each call returns 24 action rows — 1.6 seconds of motion at 15 Hz — in about 0.55 seconds on one H100.

Here is the checklist I would run before touching a real robot with it.

1. Pick the right checkpoint

Three ship on Hugging Face: the base mid-trained model, one fine-tuned on DROID, and one fine-tuned for the bimanual YAM robot. Start with the DROID fine-tune for single-arm tabletop work — it is the one behind the 46.2% number. Use the base model only if you plan your own fine-tune from scratch.

2. Size your GPU honestly

Full bf16 weights are 54.4 GB and need a single 80 GB GPU (H100/B200 class). The NVFP4 4-bit version is 17.5 GB — 68% smaller — and scores 46.5% vs 47.8% bf16 on RoboLab-120, effectively the same. That fits a 24 GB RTX 5090 or RTX PRO 6000 with room for the runtime. Runway also supports Jetson Thor for on-robot inference. My rule: prototype on the quantized checkpoint on a workstation card; reserve the 80 GB card for the final evaluation run.

3. Budget the real cost: ~$0.12 per 10-minute task

At ~$2/hour for a cloud H100, one 0.55-second chunk costs about $0.00031. A 10-minute task is 375 chunks — roughly $0.12 of inference. The model is the cheapest part of your stack. Spend the budget on cameras, calibration, and a human supervisor instead.

4. Bind your hardware before you trust the policy

Open weights do not equal plug-and-play motion. You must map your camera observations and joint state into the model's expected format, pick the action head for your embodiment, and usually fine-tune on local demonstrations. Runway ships full and LoRA fine-tuning code, with native LeRobot support and vLLM-Omni serving. The mid-training mix was ~58% egocentric human video, 20% general video, 18% real robot video — so the model knows how the world moves, but it does not know your gripper. The DROID fine-tune used only ~277 hours of successful episodes, which tells you how little data a good transfer needs — and how much your local data still matters.

5. Design for 46%, not 100%

First place at 46.2% means failing more than half the time on simulated tabletop tasks. Runway's own report names the failure modes: look-alike targets, vague instructions, long multi-step tasks. My own arithmetic: two independent attempts lift success to ~71%, three to ~84% — and retries cost a third of a cent each. Build the loop (attempt → verify → retry or re-plan), not the single shot. And remember these are Runway's own numbers with no independent reproduction yet.

6. Verify the checkpoint like code — because it moves hardware

This is the week Microsoft disclosed "Lazarus": reprogramming fewer than ten dormant neurons can plant a backdoor that passes normal evaluations and fires on a trigger. Seven altered neurons hit ≥98.8% trigger effectiveness in their test. A robot policy is a supply-chain artifact. Pin the publisher, checkpoint hash, and version; diff against a known-good baseline; audit the checkpoint; simulate first; and never run a fresh download unsupervised on a real robot. Emergency stops are a feature.

7. Test on your embodiment, not the leaderboard

Praxis-1's most interesting number is not 46.2% — it is 80.0% on SimplerEnv after fine-tuning on Google's unseen RT-1, versus 52.4% for the original RT-1. The video-learned knowledge transfers to new hardware. That means your evaluation should be on your robot and your tasks, in simulation first, then staged real-robot trials with an operator present.

The bottom line

For the first time, a competitive robot foundation model is a download, not an API key. The moat moves from "who has a model" to "who has the best demonstrations, the safest loop, and the most reliable hardware integration." That is a better competition — and it starts with a checklist, not a press release.

Sources: Runway, "Praxis-1 Is Now Available" (Nov 11, 2026); SXF Brief Issue 012 (Nov 12, 2026); AlphaSignal (Nov 11, 2026); Microsoft Research "Lazarus" disclosure (Nov 10, 2026).