Cosmos3-Nano-LIBERO10-SFT

A community fine-tune of NVIDIA Cosmos3-Nano for LIBERO-10, also called LIBERO-Long. The model jointly predicts robot action chunks and future video from camera observations and a task instruction. This is not an official NVIDIA checkpoint.

This repository contains the frozen SFT policy used in our LIBERO evaluations. No optimism-based reinforcement learning or correction training has been applied to these weights.

Download

hf download vecum0814/Cosmos3-Nano-LIBERO10-SFT --local-dir ./Cosmos3-Nano-LIBERO10-SFT

The seven BF16 model shards are approximately 31.5 GB. The included Wan2.2 video VAE adds approximately 2.82 GB. Both the Qwen vision encoder and the tokenizer/processor are already included. No separate Qwen language-model checkpoint is required for the supplied inference entry point.

Run the policy

Use Linux, an NVIDIA GPU, and a compatible Cosmos Framework environment. The release uses:

• Framework commit: de925b6b4af7f7a0f1ae85495e19f58bf2b43f39
• Recorded evaluation environment: PyTorch 2.10.0+cu130, Transformers 4.57.6, Cosmos Framework 1.2.2.
• The release smoke test uses one NVIDIA B300. A minimum GPU memory requirement has not been benchmarked.

Set up the pinned framework using its installation instructions:

git clone https://github.com/NVIDIA/cosmos-framework.git
cd cosmos-framework
git checkout de925b6b4af7f7a0f1ae85495e19f58bf2b43f39
# Install the dependencies following docs/setup.md for your CUDA environment.

With that environment active, start the included server wrapper. Replace the model directory with the directory downloaded above:

python /path/to/Cosmos3-Nano-LIBERO10-SFT/serve_libero.py --gpu 0 --port 8000 --seed 1

The wrapper resolves the VAE and processor paths relative to the downloaded model, uses the included action statistics, and matches the evaluation's action-only server configuration. It does not require editing the downloaded weights or installing our research-specific noise patches. It disables the optional prompt/video guardrail runners, matching the evaluation configuration. The default HTTP bind address is 127.0.0.1; use --host explicitly for another binding.

Use the native LIBERO client from the same framework revision. The policy interface uses:

• Two RGB camera views: third-person and wrist, each supplied at 256 × 256.
• 16-action chunks and 17-frame video predictions (including the conditioning frame).
• 10-dimensional native actions: translation (3), 6D rotation (6), and gripper (1).
• frame_wise_relative actions, quantile_rot normalization, and the included action_stats.json.
• UniPC, 30 denoising steps, guidance 1.0, and FPS 20.

The native client performs the conversion from the policy's rotation representation to simulator controls. Do not directly send the 10-dimensional output to a simulator expecting a different action space. This is a Cosmos world-action checkpoint, not a generic Transformers text-generation model.

Training provenance

• Initialization: pretrained nvidia/Cosmos3-Nano (PT-init).
• Dataset: the libero_10 subset of nvidia/LIBERO_LeRobot_v3; recorded training size 94,250 samples.
• Task adaptation: LIBERO-10/Long only, not LIBERO-Plus or all LIBERO suites.
• Recipe: action_policy_libero_10_nano.toml, preset A; global batch size 2,048.
• Weights: EMA, exported as BF16 safetensors. All seven weight shards are unchanged from the evaluated checkpoint.
• Internal source identifier: Cosmos3-Nano-LIBERO10-SFT-r2. The published repository omits the internal run suffix.

Training deviation: the trunk received 2,000 iterations, while the action head received 1,000. The second training run warm-started from the first run's weights, but the configuration reinitialized the action head and restarted AdamW and the learning-rate schedule. Training used four ranks with per-rank batch 64 and accumulation 8. This is not an exact reproduction of the reference's uninterrupted 2,000-iteration training procedure.

Evaluation

Evaluation	Successes / episodes	Success rate
LIBERO-10 / Long, 10 tasks × 50 trials	476 / 500	95.2%
LIBERO-Plus variants derived from LIBERO-10 / Long	1,840 / 2,519	73.04%

The Plus result covers the Long-derived subset, not the entire LIBERO-Plus benchmark. It uses each variant's first initial state, simulator seed 0, policy seed 1, and no intervention. The horizon is 520 simulator steps including 10 warmup steps and at most 510 policy actions. Official task instructions were retained, including generated suffixes in non-language perturbations. Both recorded evaluations completed with zero technical errors.

The original LIBERO-10 score numerically matches the published Cosmos3 PT-init result. The training deviation above prevents claiming an exact training reproduction. evaluation.json records aggregate results and source-summary hashes. validation.json describes checks performed on the publication package; a release smoke test is not a new benchmark evaluation or a clean-environment installation test.

Included components and provenance correction

• Vision encoder: 351 tensors, totaling 576,388,336 parameters, are already merged into the seventh model shard under model.net.language_model.visual.*. Their source is Qwen/Qwen3-VL-8B-Instruct at revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b.
• Tokenizer/processor: included from the same Qwen revision.
• Video VAE: vae/Wan2.2_VAE.pth, from Wan-AI/Wan2.2-TI2V-5B at revision 921dbaf3f1674a56f47e83fb80a34bac8a8f203e, verified against the upstream SHA-256.

The original export sidecar incorrectly marked the vision source as bundled: false, and the old research README incorrectly described zero vision tensors. Inspection of every safetensors header confirms that the vision weights were present. This release corrects that metadata without changing any learned weights. It also replaces the server-specific VAE path with a portable path and removes an unused internal pretrained-model path. See release_provenance.json and SHA256SUMS.

License and attribution

Cosmos3-derived model materials and the included NVIDIA normalization data are distributed under OpenMDW-1.1, retaining NVIDIA's notice of origin. The included Qwen components and Wan2.2 VAE retain their Apache-2.0 terms; see licenses/Apache-2.0.txt and NOTICE.

Upstream projects: NVIDIA Cosmos, Cosmos Framework, Qwen3-VL, Wan2.2, LIBERO, and LIBERO-Plus.