Cosmos3 Nano Policy DROID Marlin W4A16 AttnW8

Base model | Cosmos Lite | Benchmark | Cosmos 3 report

Deprecated / not recommended for new deployments. This checkpoint remains available for reproducibility of earlier experiments. Use Cosmos3-Nano-Policy-DROID-GenW8A8 instead.

This is a historical Cosmos Lite W4-AttnW8 artifact for the Nano DROID policy. It is not a recommended new deployment target.

At A Glance

Quant.	Denoise	Guidance	Peak VRAM	Request latency	Banana SR
W4-AttnW8	2	3.0	16.21 GB	4,153 ms median	84% (42/50)

The preview above is a historical RoboLab Banana rollout associated with this artifact.

Model Overview

nvidia/Cosmos3-Nano-Policy-DROID is a 16B Cosmos3 Mixture-of-Transformers policy that predicts DROID robot actions from language, RGB observations, and robot state. Cosmos Lite changes its inference representation and runtime; it does not retrain or distill the policy.

This historical attention_w8 artifact uses W4-AttnW8 across 504 (216 / 288 W4 / W8). Activations and non-targeted weights remain BF16. The bundle is self-contained and does not require the BF16 source checkpoint at deployment time.

Calibration And Evaluation

No additional policy training is performed. W4 scales use activation 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. Closed-loop quality in this card is the historical 50-episode Banana comparison.

Policy Interface

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.

Benchmark Results

This artifact is retained for reproducibility of the earlier attention_w8 experiment. The historical Banana result above is a 50-episode paired task-local comparison, not a substitute for the RoboLab-120 release benchmark. See the legacy ablation tables for the complete comparison and the main benchmark for recommended artifacts.

The artifact uses the Cosmos Lite packed weight layout and must be loaded through the Cosmos Lite runtime.

Quickstart

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

python -m pip install -U "huggingface_hub[cli]"
hf download "XXXXyu/Cosmos3-Nano-Policy-DROID-Marlin-W4A16-AttnW8" --local-dir /data/cosmos_lite/nano_marlin_w4a16_attnw8

git clone https://github.com/xxxxyu/cosmos-lite.git
cd cosmos-lite
CUDA_VISIBLE_DEVICES=0 examples/robolab_quant/pipeline.sh setup --with-sage

Validate the bundle before loading it:

BUNDLE_DIR=/data/cosmos_lite/nano_marlin_w4a16_attnw8 \
STRATEGY=attention_w8 \
examples/robolab_quant/pipeline.sh validate

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

Quantization And Bundle Details

Field	Value
Strategy	`attention_w8`
Packed linear modules	504 (216 / 288 W4 / W8)
Bundle size	15.18 GB
Activation quantization	None; activations remain BF16

Format Note

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.

Provenance

Base policy: nvidia/Cosmos3-Nano-Policy-DROID
Runtime: Cosmos Lite
Bundle contents include runtime configuration, source revisions, file sizes, and SHA256 hashes.

Limitations And Safety

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.