Cosmos Lite v0.2.0

Cosmos Lite v0.2.0

This release adds end-to-end quantized deployment support for the 4B NVIDIA
Cosmos3 Edge DROID robot policy on a single RTX 4090, alongside the existing
Cosmos3 Nano and RoboCasa365 pipelines.

Highlights

• Packed W4A16, W8A16, and fixed mixed-precision Marlin inference for
  Cosmos3 Edge.
• Streaming export and self-contained Edge bundles with the native processor,
  vision encoder, VAE, provenance, sizes, and SHA256 integrity metadata.
• Direct bundle validation, serving, replay, latency profiling, and RoboLab
  closed-loop evaluation.
• Four prebuilt Edge DROID variants with model cards, rollout videos, and
  reproducible RTX 4090 benchmarks.
• A validated two-step sampler for Edge, selected from paired closed-loop
  comparisons rather than latency alone.

Recommended Configurations

Environment and task	Quantization	Denoise	VRAM	Request latency	Success rate
RoboLab Banana	W8A16	2	8.71 GB	576 ms	72%
RoboLab Banana	W4A16-GenW8	2	8.79 GB	570 ms	80%

Both rows use guidance 3. Latency and VRAM are RTX 4090 measurements; request
latency excludes RoboLab and IsaacSim environment work. Success rates use 50
paired rollouts. See the benchmark document for the complete four-model
matrix, BF16 references, statistical tests, and protocol details.

Models

Prebuilt Cosmos3 Nano and Edge DROID policy bundles are available in the
Cosmos Lite Hugging Face Collection.

Quick Start

git clone https://github.com/xxxxyu/cosmos-lite.git
cd cosmos-lite
CUDA_VISIBLE_DEVICES=0 examples/quantized_robot_policy/setup.sh

• Cosmos3 Edge benchmark
• RoboLab pipeline
• RoboCasa365 pipeline
• Runtime guide

Cosmos Lite is community maintained and is not an NVIDIA product. NVIDIA
Cosmos Framework remains subject to its upstream license and notices; model
weights may have additional terms at their download locations.

Cosmos Lite v0.1.0

Cosmos Lite v0.1.0

The first public release of Cosmos Lite: an unofficial deployment extension
for running 16B NVIDIA Cosmos 3 robot policies on a single 24GB RTX 4090.

Highlights

• Packed W4A16, W8A16, and fixed mixed-precision Marlin inference.
• Direct loading below 24GB on the tested RTX 4090 configurations.
• Streaming quantization from BF16 or DCP checkpoints without loading the full
  source model onto the GPU.
• Self-contained bundles with config, tokenizer, VAE, provenance, sizes, and
  SHA256 integrity metadata.
• Reproducible replay, latency, and closed-loop rollout pipelines for RoboLab
  and RoboCasa365.
• Locked CUDA 12.8 policy runtime isolated from simulator dependencies.

Recommended Configurations

Environment and task	Quantization	Denoise	VRAM	Request latency	Success rate
RoboLab Banana	W8A16	2	21.42 GB	2,403 ms	90%
RoboLab Banana	W4A16-GenW8	2	18.03 GB	2,433 ms	100%
RoboCasa365 CloseFridge	W4A16-AttnW8	4	14.28 GB	1,231 ms	96%

Latency and VRAM are RTX 4090 measurements. See the repository benchmark
documents for complete protocols, sample counts, and qualifications.

Models

Prebuilt DROID policy bundles are available in the
Cosmos Lite Hugging Face Collection.

Quick Start

git clone https://github.com/xxxxyu/cosmos-lite.git
cd cosmos-lite
CUDA_VISIBLE_DEVICES=0 examples/quantized_robot_policy/setup.sh

• RoboLab pipeline
• RoboCasa365 pipeline
• Runtime guide

Cosmos Lite is community maintained and is not an NVIDIA product. NVIDIA
Cosmos Framework remains subject to its upstream license and notices; model
weights may have additional terms at their download locations.