Infernux · 熔炉

An experimental general-purpose game engine with a C++/GPU runtime and Python as a first-class development interface.

简体中文 · Website · Documentation · Forum · Releases

The image above is a real editor capture from the 0.3.4 showcase: 65,536 ordinary GameObjects share a mesh and material while a programmable RenderStack applies lighting, fog, color treatment, tilt-shift depth of field, and MSAA.

Infernux is being built as a general-purpose game engine, not as an AI wrapper around an existing editor. Performance-sensitive runtime work lives in C++17 and Vulkan. Gameplay, components, editor extensions, content workflows, automation, and render orchestration use Python 3.12 through pybind11.

The long-term thesis is that an engine for the AI era needs more than a chat panel. Models and agents need explicit access to engine semantics, world state, high-throughput data, reproducible execution, and governed tools. Infernux has some useful foundations for that direction today, but most of that system is still planned work. The project does not claim otherwise.

The current public release is a Windows-first development baseline. It can be used to author and package games on Windows x64, and it establishes a broad but young set of traditional engine tools. Version 0.3.6 unifies scene and UI objects in one Hierarchy, restores update discovery in packaged Hub builds, and publishes independently installable full Hub packages.

Android, Web, and Linux Player builds are not available in 0.3.6. Infernux also does not currently provide a general PyTorch Vulkan backend, arbitrary Python-package compatibility in exported mobile/Web games, or a finished VLM/ LLM benchmark platform.

The intended boundary is simple: performance-sensitive state belongs to the native runtime, while the public production workflow remains programmable from Python. Stable identities, generation checks, document ownership, and explicit transactions are used where arbitrary Python object lifetime would be unsafe.

Current subsystems include Vulkan Forward/Forward+/Deferred rendering, PBR, RenderGraph and programmable RenderStack paths; Jolt physics; GUID-based asset records and import artifacts; 2D/3D animation foundations; runtime Canvas UI; GPU Particle Graph; prefabs; hot reload; and a multi-panel editor with shared command, document, save, and undo/redo services.

The road from 0.3.6 to 0.5.2 is organized around three connected delivery tracks:

1. Ship games beyond Windows. Android and Web Player packages are the first product gate; Linux and headless rendering follow for deployment, CI, and simulation. Cross-platform Players may omit Numba and use an explicitly documented portable Python subset.
2. Connect the existing ML ecosystem. The Host/Editor should use standard PyTorch for development, training, and export. Shipped Players should carry only a suitable inference runtime and model, not a permanent CUDA Toolkit or the complete training stack. Android and Web providers will be selected from measured integration spikes rather than promised in advance.
3. Make worlds and tools machine-usable. Authoritative schemas, snapshot/replay, batched worlds, Torch/DLPack/buffer exchange, and formal Engine Tools are meant to serve editor UI, Python users, automation, and agents from the same semantics instead of accumulating parallel APIs.

Version 0.5.2 is the convergence point for this development cycle, not 1.0 and not a claim that Infernux already replaces Unity, Unreal Engine, or Godot. See the public roadmap for milestones.

For most users, download the current Windows x64 installer from GitHub Releases and let InfernuxHub install and manage engine versions.

To build from source, use Windows 10/11 x64, Python 3.12+, Vulkan SDK 1.3+, CMake 3.22+, Visual Studio 2022, and the MSVC v143 toolset:

git clone --recurse-submodules https://github.com/ChenlizheMe/Infernux.git
cd Infernux
conda create -n infernux python=3.12 -y
conda activate infernux
pip install -r requirements.txt
cmake --preset release
cmake --build --preset release

Launch the Hub from source:

python packaging/launcher.py

Run Python and native test suites from the repository root:

python -m pytest python/test/ -v
cmake --preset native-tests
cmake --build --preset native-tests
ctest --preset native-tests --output-on-failure

• Documentation hub
• API reference
• Release notes
• Contributing guide
• Repository automation
• Technical report

Repository-level commands live under scripts/; website-only deterministic generators and contract tests live under docs/tools/. Build and package output goes to out/, while final local release artifacts go to dist/releases/<version>/.

Regenerate the checked-in API documentation intentionally with:

scripts\docs\update_api_docs.bat

@software{chen2026infernux,
  author  = {Chen, Lizhe},
  title   = {Infernux},
  year    = {2026},
  version = {0.3.6},
  url     = {https://github.com/ChenlizheMe/Infernux}
}

Bug reports, feature requests, reproducible tests, platform bring-up work, and workflow feedback are welcome. Because the engine is still young, include the exact engine version, environment, reproduction steps, and affected layer in an issue. Read CONTRIBUTING.md, SECURITY.md, and SUPPORT.md before submitting sensitive or large changes.

Infernux is released under the MIT License.