Run NVIDIA's 16B Cosmos3 Nano DROID robot policy on a single 24GB RTX 4090. This self-contained GenW8A8 deployment is the recommended Cosmos Lite tradeoff for Nano.

The release protocol is one RTX 4090, batch size one, guidance 3, two UniPC denoise steps, shift 5, and a 32-action chunk. The result is 380/1,200 closed-loop rollouts across 120 RoboLab tasks, 10 episodes per task, using default instructions. Request latency excludes simulator time.

A representative Nano GenW8A8 closed-loop RoboLab episode is shown above. It uses the validated guidance 3 / two-step sampler.

Cosmos 3 is a family of omnimodal world models for Physical AI. The 16B Cosmos3-Nano model uses a Mixture-of-Transformers (MoT) architecture with an autoregressive transformer for discrete tokens and a diffusion transformer for continuous modalities. This DROID policy predicts robot actions from language, RGB observations, and robot state. Cosmos Lite changes the inference representation and runtime; it does not retrain or distill the policy.

GenW8A8 applies dynamic per-token FP8 E4M3 W8A8 to 252 action-generation linear modules. The other 252 selected language/MoT linear modules use packed W4A16; activations outside the W8A8 modules remain BF16. The bundle is self-contained and does not require the BF16 source checkpoint or calibration dataset at deployment time.

No additional policy training is performed. W4 scales use input-channel statistics from 128 successful episodes in the official nvidia/Cosmos3-DROID training split. Calibration does not update model parameters and uses no RoboLab evaluation episode. FP8 activation scales are computed dynamically per token at inference time.

The public result above covers all 120 RoboLab tasks. Full protocol, confidence intervals, topology, and memory definitions are in the Cosmos Lite benchmark.

The validated RoboLab client composes three camera views into a 640x540 RGB input and maps it to the model's 736x544 inference bucket. It returns a 32x8 DROID joint-position action chunk; the standard integration executes all 32 actions before requesting another chunk.

Model precision is fixed by this artifact. Guidance, denoise steps, and shift are independent runtime controls. The release config defaults to guidance 3, two denoise steps, and shift 5.

The primary release result is the full RoboLab-120 evaluation above. It is compared with Nano W8A16 in the main benchmark; sampler and legacy quantization experiments are separated into RoboLab ablations. The recommended role of this artifact is the Nano latency/memory tradeoff; Nano W8A16 remains the calibration-free alternative.

Install the Hugging Face CLI, clone the runtime, and download this self-contained bundle:

Validate the bundle before loading it:

For serving and closed-loop RoboLab evaluation, follow the RoboLab pipeline guide.

This is a Cosmos Lite deployment bundle, not a generic GPTQ or AWQ checkpoint. The packed weights, precision map, residual tensors, processor assets, and runtime manifest must be loaded through Cosmos Lite. Weight-only artifacts keep activations in BF16; GenW8A8 artifacts use FP8 W8A8 only in the generation branch and retain the documented BF16 paths elsewhere.

The default server binds to 127.0.0.1:8000 and has no built-in TLS or authentication. Do not expose it directly to an untrusted network.

Simulation results do not establish transfer to a new robot, camera layout, task distribution, or action contract. Validate those changes with paired closed-loop rollouts. This policy is not real-robot safety certified. Real deployment requires an independent E-stop, watchdog, motion limits, stale-command rejection, and operator supervision.

This community quantization is not an NVIDIA product.

The upstream model and this derivative are distributed under OpenMDW-1.1. Review the upstream model card and repository notices before use.