Run NVIDIA's 16B Cosmos3 Nano DROID robot policy on a single 24GB RTX 4090. This self-contained W8A16 deployment is the calibration-free Cosmos Lite option 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 378/1,200 closed-loop rollouts across 120 RoboLab tasks, 10 episodes per task, using default instructions. Request latency excludes simulator time.

This historical Banana rollout preview uses the same W8A16 artifact with guidance 3 and four denoise steps. The RoboLab-120 result above uses the current two-step release 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.

This artifact packs all 504 selected language/MoT linear modules to W8. Activations and non-targeted weights remain BF16. The self-contained bundle uses weight statistics only, so it needs no calibration dataset and does not require the BF16 source checkpoint at deployment time.

No additional policy training is performed. Full W8A16 is calibration-free: packing uses weight statistics only and does not consume rollout or training observations.

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 GenW8A8 in the main benchmark; sampler and legacy quantization experiments are separated into RoboLab ablations. Use this artifact when calibration-free deployment is required; use Nano GenW8A8 for the recommended latency/memory tradeoff.

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.