Open-weight world models
Every downloadable world model we have verified: weights actually downloadable, license as stated on the model card, checked 2026-07-18.
| Model | Org | Params | License | Branch | Modality |
|---|---|---|---|---|---|
| Cosmos3-Super | NVIDIA | 64B | OpenMDW-1.1 | Agents & robotics | Omnimodal in/out: text, image, video, audio, action |
| Cosmos3-Nano | NVIDIA | 16B | OpenMDW-1.1 | Agents & robotics | Omnimodal in/out |
| Cosmos3-Super-Text2Image / -Image2Video | NVIDIA | 64B | OpenMDW-1.1 | Video-generative | text→image / image→video |
| Cosmos3-Nano-Policy-DROID | NVIDIA | 16B | OpenMDW-1.1 | Agents & robotics | Vision-language → robot actions |
| V-JEPA 2 | Meta FAIR | 0.3B–1B (8 checkpoints) | MIT | Latent prediction | Video → latent representations (no pixel generation) |
| Oasis 500M | Decart + Etched | ~500M | MIT | Game / simulation | Keyboard/mouse actions → Minecraft gameplay video |
| DIAMOND | Alonso et al. (U. Geneva / Edinburgh) | 381M (CS:GO); Atari agents | MIT (code; HF weights untagged) | Game / simulation | Actions → playable Atari / CS:GO gameplay (diffusion world model) |
| WHAM (Muse) 200M / 1.6B | Microsoft Research | 200M / 1.6B | Microsoft Research License (research-only, NOT open source) | Game / simulation | Bleeding Edge gameplay: frames + controller actions → gameplay continuation |
| Matrix-Game 1.0 | Skywork | 17B | MIT | Game / simulation | Image + keyboard/mouse → Minecraft world video |
| Matrix-Game 2.0 | Skywork | 1.8B | MIT | Game / simulation | Image + frame-level actions → real-time 25 FPS streaming video |
| Vista | OpenDriveLab | SVD fine-tune (size not verified) | Apache-2.0 | Driving | Driving video + actions (steering, speed, trajectory, goal) → future driving video |
| DreamerV3 | Google DeepMind (Hafner) | code + configs S–XL | MIT | Agents & robotics | Pixels + actions → control (RSSM world model) |
Notes
- Cosmos3-Super — Mixture-of-Transformers: 32B reasoner + 32B generator; ~133 GB BF16. Released 2026-05-31.
- Cosmos3-Nano — The fine-tunable small sibling — realistic on a multi-GPU RunPod budget.
- Cosmos3-Super-Text2Image / -Image2Video — Generation-specialist variants of Super.
- Cosmos3-Nano-Policy-DROID — Robot policy fine-tuned on DROID.
- V-JEPA 2 — ViT-L/H/g backbones at 256/384 + SSv2/Diving48 fine-tunes; action-conditioned vjepa2-ac ships separately. arXiv:2506.09985.
- Oasis 500M — Downscaled from the (unreleased) demo model; DiT + ViT-VAE, Diffusion Forcing. Inference code: github.com/etched-ai/open-oasis. Oasis 2/3 are hosted-only.
- DIAMOND — NeurIPS 2024 Spotlight; 1.46 mean HNS on Atari 100k (best at publication for in-model-trained agents); CS:GO model trained 12 days on one RTX 4090. arXiv:2405.12399.
- WHAM (Muse) 200M / 1.6B — Nature Feb 2025; 1.6B causal transformer + ViT-VQGAN tokenizer; trained on 98 H100s × 5 days. Downloadable but non-commercial.
- Matrix-Game 1.0 — May 2025; inference needs ≥80 GB GPU memory. arXiv:2506.18701.
- Matrix-Game 2.0 — Aug 2025; three distilled checkpoints (universal, GTA driving, TempleRun); runs on a single 24 GB GPU. arXiv:2508.13009.
- Vista — NeurIPS 2024; weights on HF since June 2024; fine-tuned from Stable Video Diffusion; author-demonstrated use as a reward function. arXiv:2405.17398.
- DreamerV3 — Nature 2025 (640:647–653). Train-it-yourself algorithm rather than a released checkpoint; the standard MBRL baseline.
In verification: Large World Model (LWM), Hunyuan-GameCraft, Decart Mirage, and the Genie open reproductions. Entries are added only after the weights, license, and claims check out. Note the license column carefully — "downloadable" and "open source" are not the same thing (see WHAM).