Oasis is not a conventional Minecraft clone. It is an experimental, action-conditioned world model from Decart and Etched that generates Minecraft-like frames in real time. The public release includes inference code and weights for a smaller 500-million-parameter model, but not a traditional voxel engine with editable physics, world files, and deterministic game rules.
The project was covered by Hackaday on November 9, 2024: Here’s Code For That AI-Generated Minecraft Clone. The most useful way to understand the headline is: Oasis makes gameplay appear through neural video generation, rather than using AI to write an ordinary Minecraft replacement.
What Oasis actually is
The official Oasis project describes an “experiential, realtime, open-world AI model.” It accepts keyboard actions and recent visual context, then generates the next part of the experience. The demonstrations show movement, jumping, block breaking, item pickup, building, lighting, inventory interactions, animals and tool-dependent behavior.
There is no conventional physics engine underneath those visuals. Apparent physics, rules, graphics and interactions are produced by the model. That makes Oasis a research prototype exploring whether a generative model can act as a game engine, not an unofficial port of Minecraft Java or Bedrock.
#1 Best Overall
- Collectible figurine inside the grip and a multi-layered design featuring iconic game scenes
- Built-in, rechargeable battery and a 30-foot wireless range
- Expertly designed for all-day comfort and long-lasting performance
- Brighten your gameplay with LED lighting effects that you can change using the ABXY buttons
- Officially licensed by Nintendo and Mojang Studios
Is it really a Minecraft clone?
“Minecraft-like” is the accurate description. Oasis uses Minecraft-style visual and gameplay patterns, but it does not reproduce Minecraft’s codebase, world format or complete ruleset. It is not affiliated with Mojang or Microsoft, and it should not be presented as a persistent replacement for Minecraft.
The word clone also hides an important technical difference: a conventional clone stores blocks and entities as explicit data and renders that state. Oasis predicts what the next frame should look like after an action. A scene can therefore look convincing without having a stable, inspectable world state behind it.
What “AI-generated Minecraft” can mean
Several very different technologies are routinely described with the same phrase.
AI-generated game code
An AI coding tool can write Rust, JavaScript or engine scripts for a normal voxel game. The resulting program still uses ordinary rendering, physics, networking, procedural generation and save files. Minecraft: Vibed Edition, for example, is described by its author as having approximately 99% AI-generated code, with manually made textures and human debugging. Its source is available at GitHub.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Rank #2
- Paying tribute to the iconic Creeper, this Minecraft Gaming Locker is the ultimate storage solution for gamers.
- This universal vertical gaming locker boasts the ability to hold up to 10 of your favourite games or Blu-Ray films. Hang up to 4 gaming controllers on the side of the locker
- Give your gaming headset the perfect home as this gaming locker includes a secure headphone holder on the top to save space
- Keep your gaming space tidy with extra compartments for all your gaming essentials. Store all your accessories, cables, remotes and other equipment securely in the bottom drawer
- Official Minecraft merchandise, designed and engineered by Numskull Designs
AI-generated assets
AI can make textures, models, sounds or other content while a conventional engine runs the game.
AI-generated gameplay frames
Oasis belongs primarily in this category. The neural model generates the visible result frame by frame, conditioned on player actions. It is not primarily an LLM writing a voxel engine.
A later browser project, Vibe Craft, illustrates the other category: its creator describes Three.js/WebGL, Node.js, WebSockets, chunk streaming, persistence, inventory, crafting, mobs and server-authoritative state, with code generated by ChatGPT Codex 5.3. Oasis asks whether a model can be the runtime; AI-coded projects ask whether an AI can build a conventional runtime.
Where the code and weights are
The public repository is etched-ai/open-oasis. It includes model architecture and inference code, checkpoint-loading logic, generation scripts and weights for the downscaled Oasis 500M model. The repository displays an MIT license.
Rank #3
- MODERNIZED DESIGN — Experience the modernized design of the XBOX Wireless Controller with sculpted surfaces and updated geometry that enhances comfort and control during long gaming sessions.
- PRECISION PERFORMANCE — Stay on target with a hybrid D-pad and textured grips on triggers, bumpers, and back case for improved accuracy and handling.
- SHARE BUTTON: Seamlessly capture and share content such as screenshots, recordings, and more with the new Share button.
- VERSATILE CONNECTIVITY — Connect via USB-C for plug-and-play on console and PC, or quickly pair and switch between supported devices with XBOX Wireless and Bluetooth support.
- BUILT-IN AUDIO SUPPORT — Plug in compatible headsets using the 3.5mm audio jack for direct voice chat and immersive in-game sound.
That does not mean you are downloading a complete, editable Minecraft implementation. The repository does not expose a normal block-collision system, structured world database, deterministic physics engine, standard multiplayer server or conventional save/load format. Much of the behavior is distributed across neural-network weights, conditioning and the inference pipeline.
The project page links the repository, model information and a live demonstration. The hosted demonstration and the local 500M checkpoint should not be assumed to have identical quality, speed or behavior.
How to run the released local model
The documented workflow is aimed at a CUDA-capable PyTorch environment, not a lightweight browser installation. Commands in the repository include:
git clone https://github.com/etched-ai/open-oasis.gitandcd open-oasis.- Install PyTorch and torchvision for the relevant CUDA environment:
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121. - Install the remaining dependencies:
pip install einops diffusers timm av. - Authenticate with Hugging Face using
huggingface-cli login. - Download the two checkpoints:
huggingface-cli download Etched/oasis-500m oasis500m.safetensorshuggingface-cli download Etched/oasis-500m vit-l-20.safetensors - Run
python generate.py, or provide explicit checkpoint paths with--oasis-ckptand--vae-ckpt. - For an image prompt, use
python generate.py --prompt-path <path to .png, .jpg, or .jpeg>.
The documented output is video.mp4. Check the repository’s current README before installing: CUDA, Python, PyTorch, dependency and checkpoint requirements can change after the 2024 release.
Rank #4
- Lenticular Minecraft Dungeons II design that shifts from Valorie to Creeper
- TMR thumbsticks deliver smooth, accurate gameplay
- Up to 30-foot low latency wireless connection
- Rechargeable battery with up to 40 hours per charge
- Dual mappable quick-action back buttons
- A compatible NVIDIA/CUDA setup is likely required by this documented path.
- Hugging Face authentication is required to obtain the checkpoints.
- No verified minimum GPU or VRAM figure is established here, so do not treat the model as suitable for any particular graphics card without testing.
- Local inference carries hardware, electricity and setup costs even though the code is publicly available.
How the model generates a world
Oasis combines a spatial autoencoder with a latent diffusion backbone; both are transformer-based, according to the official technical description. Generation is autoregressive:
- The player supplies an action such as moving, looking, jumping or interacting.
- The model receives that action and visual context.
- It predicts the next frame or continuation.
- The generated result becomes context for subsequent predictions.
- Small mistakes can accumulate as the sequence continues.
This differs from a conventional voxel game, where a physics system updates explicit coordinates, an inventory is structured data, and a renderer draws known geometry. In Oasis, the visible frame is the primary output, so plausible appearance does not guarantee exact internal state.
What it can do—and what that claim means
The project demonstrations show an impressive range: walking through terrain, jumping, breaking and placing blocks, picking up items, manipulating an inventory, observing lighting changes, interacting with animals and using tools in apparently different ways. These are demonstrations of generated behavior, not a guarantee that every interaction remains reliable in a long session.
The project reports generation at 20 frames per second, or roughly one frame every 0.04 seconds. That is a project-reported figure dependent on the inference stack and hardware; it does not establish equal performance for the local 500M release or prove low latency, temporal stability or game-quality responsiveness. The significance is that interactive generation requires far more throughput than offline text-to-video systems, which the project contrasts with systems taking 10–20 seconds to make one second of video.
Best Value
- Customizable RGB lighting illuminates the iconic Minecraft design
- Wireless play for XBOX and Windows 10/11 PCs via USB transmitter
- Bluetooth (BLE) connection for compatible Android mobile and streaming devices
- Ultra-sensitive TMR thumbsticks for lighting-fast control & precision
- Hall-Effect hair triggers with 3-way trigger stops for faster response and greater control
Why the experience breaks down
The same official page that presents Oasis’s capabilities lists substantial limitations:
- Temporal inconsistency: objects and terrain can change between frames.
- Limited long-horizon memory: the model may lose track of earlier actions and locations.
- World drift: built areas and previously seen details may not persist exactly.
- Imprecise inventory and object control: items, quantities, blocks and tools may behave inconsistently.
- Fuzzy distance rendering: faraway or detailed scenes can lose structure.
- Domain-generalization problems: unfamiliar situations are more likely to produce errors.
- Autoregressive error accumulation: a small prediction mistake can compound over a long sequence.
Hackaday’s account also describes coherence degrading after prolonged attention to an object or movement through dark and repetitive areas. In practical terms, Oasis can produce a compelling moment of play while failing the central test of a persistent game: remembering exactly what exists and why.
Is Oasis playable?
- As a technical demonstration: yes. It shows action-conditioned, real-time generative video in an interactive setting.
- As a conventional game: only superficially. The demonstrations resemble gameplay, but reliability, persistence and precise control are limited.
- As a Minecraft replacement: no. It is neither an official Minecraft implementation nor a stable substitute for survival, building or multiplayer play.
- As a research artifact: very much so. It is a useful example of neural rendering, diffusion transformers and world-model inference.
Oasis compared with AI-coded voxel games
| Project type | What the AI generates | What runs the experience |
|---|---|---|
| Oasis | Frames, apparent rules and visual state | A neural world model and inference pipeline |
| AI-coded voxel game | Source code and possibly assets | A conventional engine and program logic |
| AI-assisted development | Individual systems, shaders, tests or fixes | Human-designed software architecture plus an ordinary runtime |
If you want stable saves, multiplayer, mod support, deterministic physics or a codebase that can be extended feature by feature, use a conventional stack such as Godot, Unity, Unreal, Three.js/WebGL, Rust with wgpu or Bevy, or an open voxel platform such as Luanti. AI coding tools can accelerate that work, but they do not remove the need for architecture, testing and debugging.
Licensing and practical cautions
The repository’s MIT license covers the released code under its terms; it does not automatically answer questions about training data, generated outputs, Minecraft trademarks, similar-looking assets, redistribution of modified models or third-party dependencies. Do not treat the license as a definitive copyright or fair-use opinion.
Free tools Windows power users keep installed
One-click scans. No signup required.
Likewise, do not describe Oasis as “free” without qualification. The repository may be available without a software purchase, but local inference requires suitable hardware and electricity, while hosted access can have its own availability or usage conditions. The project is unofficial and should not be marketed as Mojang- or Microsoft-endorsed Minecraft.
Bottom line
Oasis is important because it tests a radical idea: a generative model can render an apparently interactive game world instead of merely supplying assets or writing game code. Its public 500M release lets technically capable readers inspect and run that approach. But the repository is an inference system, not a conventional Minecraft clone, and the model’s memory, consistency, control and hardware demands keep it firmly in research-demo territory.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
