Gamma-World: Multi-Agent World Model
A generative world model that simulates a shared environment for multiple independently controllable agents, scaling beyond two players.
- Source repo
- nv-tlabs/Gamma-World
- Stars
- ★ 579
- Last updated
- 5mo ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 38/100 · Major gaps
At a glance
- Works with
- Portable with changes
- You'll need
- Typical use
- Researchers studying multi-player game world models who need interactive simulated environments
- Main limitation
- Setup, inference, and training details live in docs/setup.md, docs/inference.md, and docs/training.md, which are not cached in the provided source, so exact commands cannot be confirmed
- Website
- research.nvidia.com ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-06-17
- Source review
- 38/100 · Major gaps 2 safety controls not found
What does this agent do, and when should you use it?
Gamma-World is an open-source generative multi-agent world model from NVIDIA research that rolls out a single shared environment in which multiple agents act simultaneously and remain independently controllable. It ingests synchronized observations and actions per agent, tokenizes each stream with shared visual and action encoders, and generates future multi-agent rollouts with a causal multi-agent DiT. Simplex Rotary Agent Encoding gives each agent a distinct rotary phase while keeping all agents permutation-equivalent, and Sparse Hub Attention routes cross-agent information through learnable hub tokens in linear time. A full-context diffusion teacher is distilled into a block-causal student that streams at 24 FPS with KV caching, and the encoding enables zero-shot generalization from two to four players.
The repository implements both inference and training for multi-agent video world modeling. It reads synchronized observations and actions from multiple agents, tokenizes each agent's stream with shared visual and action encoders, and runs a causal multi-agent DiT to produce future multi-agent rollouts. Inference supports bidirectional, causal, and causal-few-step models on user-supplied initial frames and actions, and the distillation stage converts the teacher into a causal student that generates temporal blocks sequentially with KV caches at 24 FPS. Training converts a released checkpoint to DCP and then runs the teacher, causal, and DMD stages.
- Researchers studying multi-player game world models who need interactive simulated environments
- Robotics labs exploring multi-robot coordination through shared world model rollouts
- Teams wanting to generalize from two-player training data to four-player scenes without retraining
- Engineers distilling a diffusion teacher into a real-time causal student
- Researchers benchmarking permutation-symmetric encodings and cross-agent attention designs
How do you install or deploy this agent?
Installation is documented in docs/setup.md, which is not included in the provided source, so exact commands cannot be given here. The runtime requires Python, PyTorch, and GPU hardware. Start by cloning the repository and then follow the setup document:
git clone https://github.com/nv-tlabs/Gamma-World.git
cd Gamma-WorldThen run the environment and dependency installation described in docs/setup.md.
How do you use this agent?
Inference is documented in docs/inference.md and covers running the bidirectional, causal, and causal-few-step models on your own initial frames and actions. Training is documented in docs/training.md and covers converting a released checkpoint to DCP and training the teacher, causal, and DMD stages. Exact commands are not included in the provided source; consult these documents.
# View the inference and training documentation
cat docs/inference.md
cat docs/training.mdThe project page provides videos and qualitative results for evaluating the model before running it.
What are this agent's strengths and limitations?
- Simplex Rotary Agent Encoding is parameter-free and permutation-symmetric, removing learned per-slot identities and a fixed agent order
- Sparse Hub Attention reduces cross-agent attention from O(N²) to O(N) and is shown to lower latency and FLOPs as agents increase
- Distilled causal student enables real-time 24 FPS block-causal streaming with KV caching
- Apache-2.0 license supports research and commercial use, and code plus training pipeline ship together
- Demonstrates transfer from virtual game settings to real-world multi-robot coordination
- Setup, inference, and training details live in docs/setup.md, docs/inference.md, and docs/training.md, which are not cached in the provided source, so exact commands cannot be confirmed
- Requires GPU hardware and heavy deep-learning dependencies; first run will not complete in minutes
- New 2026 release with an unsettled community ecosystem and unclear checkpoint availability
- No evidence of integration with chat platforms, API standards, or agent packaging conventions
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Gamma-World: Multi-Agent World Model This agent | 38 · Major gaps | — | ★ 579 | 5mo ago | Python | — |
| 4KAgent | 39 · Major gaps | CLIFree + model costs | ★ 842 | 1y ago | Python | — |
| Principia | 66 · Some gaps | CLIFree + model costs | ★ 1.1k | 1mo ago | Rich Text Format | — |
| AgileRL | 62 · Some gaps | Library / SDKFreemium | ★ 970 | today | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
This is a research world-model codebase, not an Agent product with permission boundaries. README and pyproject describe no permission model, confirmation flow, or data-flow disclosure; the only dependency is pinned cosmos-oss==0.1.0 via workspace override with no vulnerability audit or hash locking, so dependency_security is 1. Source files carry SPDX and Apache-2.0 headers, giving clear attribution (2). No rollback mechanism is evidenced (0).
The README's method claims (Simplex Rotary Agent Encoding, Sparse Hub Attention, 24 FPS) are internally consistent, and test files show backend skip conditions and tolerance assertions, indicating some engineering rigor (2). Dependency availability is gated on specific CUDA/GPU architectures (Hopper/Blackwell), unusable in ordinary environments, and no failure-message design is visible, each 1.
README targets researchers with setup/inference/training doc entry points, a clear scenario (2); but capability boundaries (supported inputs, VRAM needs, non-GPU use) are unstated and trigger conditions vague, and environment fit is Linux+CUDA only, each 1.
README is well-structured with a User Guide index (2); install notes only link docs/setup.md not included in the packet (1); package naming and structure are stable (2); examples are images/video only with no FAQ (1); known limitations are not listed (1); LICENSE is full Apache-2.0 and consistent with pyproject (3); no changelog/version history beyond two News items (1); maintenance responsibility is unclear with an unverified publisher (1).
As a research reproduction package, output usability depends on checkpoints and docs not provided (1); marginal value over single-agent world models exists but requires heavy compute (1); cost/benefit is unquantified (1).
Paper and arXiv links are traceable, but no benchmark scripts in the packet map to the paper's claims (1); tests and README claims lack cross-corroboration (1); fact and inference are mostly separated but not explicitly labeled (1).
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- This is a research codebase, not a deployable Agent product; it lacks permission, confirmation, and data-flow controls and should not be assessed as an Agent framework for safety.
- The cosmos-oss==0.1.0 dependency is workspace-overridden without hash locking or vulnerability audit, leaving supply-chain risk unknown.
- Running requires Hopper/Blackwell-class GPUs and specific CUDA versions; ordinary environments cannot reproduce it.
- Publisher identity is unverified by the FollowAgents registry; maintenance responsibility and update path are unclear.
- Performance claims such as 24 FPS and four-player generalization have no in-packet benchmark scripts to verify them.
FAQ
Is Gamma-World: Multi-Agent World Model open source?
Is Gamma-World: Multi-Agent World Model safe to use?
Is Gamma-World: Multi-Agent World Model still maintained?
Who makes Gamma-World: Multi-Agent World Model?
What is Gamma-World: Multi-Agent World Model's official website?
Thanks — we use this to improve the page.