Worlds

Checkpoints of the explorable world models of WanGP, prepared for its loaders. WanGP downloads them when a world model is first used.

File	Model	Content	Source	Licence
evoke_turbo_bf16.safetensors	EVOKE Turbo World 14B	Transformer in BF16. The modules the source keeps in FP32 at inference (time embedder, modulation tables, cross-attention norms) stay in FP32. Keys are renamed to WanGP's Wan names.	AlayaLab/Evoke-Turbo	Apache-2.0 (evoke/LICENSE)
evoke_turbo_int8_convrot.safetensors	EVOKE Turbo World 14B	The same transformer with its block linears in INT8 ConvRot	as above	Apache-2.0 (evoke/LICENSE)
evoke/vigeo1_1.safetensors	EVOKE world memory (depth)	ViGeo1.1 with the depth path only: projections in FP16, the rest in FP32. The ray, normal, mask, confidence and camera heads are removed.	pkqbajng/ViGeo1.1, code aigc3d/ViGeo	CC-BY-NC-4.0 (non-commercial)
hy_nomas_distilled_int8.safetensors	Hy-Nomas World Model 8B	The distilled transformer, INT8 (upstream file, unchanged)	tencent/Hy-Nomas	Apache-2.0 (hy_nomas/LICENSE)
hy_nomas/pixelumm_decoder_fp8.safetensors	Hy-Nomas World Model 8B	The decoder tensors of the fine-tuned PixelUMM FP8 decoder (bytes unchanged, prefix removed), without the unused DCAE-V encoder	tencent/Hy-Nomas (VAE/PixelUMM_FT_FP8)	NVIDIA One-Way Noncommercial (research and evaluation only, hy_nomas/LICENSE_PixelUMM)

Text encoders, tokenizers and the Wan2.1 VAE are shared with other WanGP models. They are hosted in DeepBeepMeep/Wan2.1 and DeepBeepMeep/Qwen_image.

Licences

Each file keeps the licence of its source.

• EVOKE: Copyright 2025 The Helios Team and The HuggingFace Team, Copyright 2026 The Evoke Team, Apache-2.0.
• ViGeo1.1: Creative Commons Attribution-NonCommercial 4.0 (https://creativecommons.org/licenses/by-nc/4.0/). Its DINOv2 layers are Copyright (c) Meta Platforms, Inc. and affiliates.
• Hy-Nomas: Copyright (C) 2026 Tencent. The full licence and third-party notices of the Tencent distribution are in hy_nomas/LICENSE.
• PixelUMM decoder: NVIDIA One-Way Noncommercial License, full text in hy_nomas/LICENSE_PixelUMM (https://huggingface.co/nvidia/PixelUMM).

EVOKE worlds need ViGeo1.1 for their memory, and Hy-Nomas worlds are decoded by PixelUMM. Both models are therefore for non-commercial use only. Users are responsible for complying with these licences.