Cosmos3-Edge-ThunderWorldFP4

Frozen ThunderWorld FP4 generation-branch weights, with 2× (default) and 1.5× repair profiles. These are the exact packed assets used in our full-corpus quality evaluation, not newly requantized weights.

This is a custom quantized-weight package, not a standalone Transformers/Diffusers checkpoint. The bundled loader requires the pinned NVIDIA base model (a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba) for unchanged AR, vision/VAE, normalization and other components. It loads the base model during preparation and replaces the selected generation linears with these assets. Package size is therefore not total deployment size or peak VRAM.

Paper/preprint · Code · Generated benchmark data

Method

ThunderWorld is training-free. A dense W4A4 base product is supplemented by selected activation-residual and weight-residual FP4 products. W0, compact W1, scales and supports are prepared offline; activations and repair selection always use the current input. The profiles preserve every base attention block and disable the second-order cross term. The budget caps padded request MMA work relative to base W4A4; it is not a speedup factor. The 2× profile includes jointly calibrated channel prescales, quantization scales and W1 supports. The 1.5× profile retains identity-prescale RTN weights with calibrated repair allocation and protected N128 weight blocks. Outlier splitting is not enabled in either frozen profile.

Profile	Packed weight files	Generation linears	Historical request bound
2x	0.812 GiB	168	2.000000×
1.5x	0.778 GiB	168	1.499957×

Bounds are recomputed for the current request. Plans are qualified for I2V, 320×192, 61 frames, 24 fps, 35 denoising steps, CFG 6, flow shift 3. Other tasks/resolutions require new budget qualification. The bundled backend targets SM120 (RTX PRO 6000 Blackwell / RTX 5090). Available VRAM must also fit the unquantized components and preparation: hardware support does not guarantee model fit.

Calibration data

Each model was calibrated independently using 16 frozen I2V scenarios: 9 RBench and 7 PAI-Bench-G, one seed each (301001–301016). Inputs are benchmark prompts and conditioning images; BF16 trajectories provide current-input operator references across four denoising phases and both CFG branches. No gradient updates or model training are used.

• RBench: DAGroup-PKU/RBench, revision 6bdccf349ff5a8f68302428351e94f34ecd62450.
• PAI-Bench-G: shi-labs/physical-ai-bench-generation, revision c36b7beb69bd7c4adcedc1431407fce98f9dfe44.
• calibration_manifest.json records exact IDs, seeds, image/pixel hashes, prompt hashes and the frozen 16 development / 32 validation / 72 initial-test scenario splits. Raw dataset images are not redistributed here.
• Later full-corpus evaluation includes previously used calibration/development scenarios; it is not a new held-out-scene generalization test.

Download and use

Use Python 3.12 and a compatible CUDA/PyTorch installation. The qualified software versions are in requirements-runtime.txt; install the matching PyTorch build for your CUDA environment first.

hf download Concyclics/Cosmos3-Edge-ThunderWorldFP4 --local-dir Cosmos3-Edge-ThunderWorldFP4
cd Cosmos3-Edge-ThunderWorldFP4
python -m pip install -r requirements-runtime.txt
python infer.py --profile 2x --verify-only

CUDA_VISIBLE_DEVICES=0 python infer.py \
  --profile 2x --image /path/to/input.jpg \
  --prompt "Describe the desired continuation of the scene." \
  --seed 12345 --output ./output-2x

The loader downloads the pinned base model if needed; use --base-model /path/to/base to reuse a local copy. Use --profile 1.5x for the lower-budget assets. --debug off and --execution graph are defaults. Outputs include MP4, original float32 NPZ and a JSON request/budget receipt. The release entry point makes no formal latency claim. The portable runtime is bundled to avoid depending on a moving GitHub checkout.

Quality and scope

The accompanying quality_summary.csv reports full RBench/PAI-Bench-G results. LPIPS/SSIM/PSNR measure agreement with same-model, same-input, same-seed BF16, not real-video ground truth. Qwen judge scores are adapted task/semantic scores, not official complete leaderboard scores or robot execution success. The 2× profile improves output consistency over plain W4A4 and the tested ARCQuant/SharQ adapters; these low-budget releases do not claim to exceed FP8 on pixel consistency or win every judge comparison. Super results extend the Edge/Nano experiments in preprint v1.

License and citation

Derived model weights retain OpenMDW-1.1. NVIDIA origin notices are in NOTICE and UPSTREAM_MODEL_CARD.md. The bundled runtime retains its source notices. Please cite:

Chen et al. ThunderWorld: Efficient World Model Generation with Selective FP4 Refinement via Native GEMMs. Preprints.org, 2026, version 1. https://www.preprints.org/manuscript/202610.0689