QwenGenesis
Artificial Life as Neural Computation
Created by @suayptalha
QwenGenesis explores deterministic artificial-life dynamics computed by compiled neural weights. The project demonstrates birth, predation, mutation, ecological collapse and evolutionary selection on a spatial cellular world.
Watch the three worlds
| Evolution | Ecological Collapse | Predator Apocalypse |
|---|---|---|
| Golden Evolution | Ecological Collapse | Predator Apocalypse |
What is included
genesis_native.safetensors: saved Genesis SwiGLU neural circuit weights, independent of the pretrained Qwen language-model backbone.qwengenesis/: deterministic world, compiled circuit, CUDA inference and cinematic renderer.qwengenesis_native/: Qwen-style SwiGLU implementation and comparison tests (when included).scripts/,tests/: reproducibility and scenario comparisons.videos/: three demonstration MP4s.
Genesis is executed by its independent neural circuit. The saved Genesis weights use Qwen-style SwiGLU gate, up and down projections. The native circuit performs ecosystem state transitions without loading a pretrained language model. The full-size FFN experiments use Qwen3.5-compatible MLP dimensions, but they are not pretrained decoder layers.
Reproduce
# Install dependencies
pip install torch numpy safetensors pillow imageio imageio-ffmpeg transformers huggingface_hub
# Download QwenGenesis repository
hf download suayptalha/QwenGenesis --local-dir QwenGenesis
# Enter repository directory
cd QwenGenesis
# Run Genesis standalone neural ecosystem
python run_checkpoint.py --checkpoint genesis_native.safetensors --steps 10 --seed 93 --size 24
# Run native FFN ecosystem scenarios (CUDA GPU required)
python -m scripts.run_native_scenarios --device cuda --only all --world-size 96 --output-root outputs_native
The comparison scripts and GPU outputs allow independent verification. The complete pretrained language-model decoder is not used by the standalone Genesis engine.
Verified results from the supplied run
- SwiGLU circuit vs reference: bit-state agreement over 3 × 100 steps (24×24), maximum reported absolute floating-point difference ~7.13×10⁻¹⁵.
- Full-sized Qwen3.5 FFN comparison on A100: 3 × 10 ticks with binary-state agreement.
- 96×96 demonstration scenarios:
predator_apocalypse(70 steps),ecological_collapse(150 steps),golden_evolution(180 steps); all reported output frames matched the reference binary states. - Illustrative predator trajectory (apocalypse): 2,795 at step 10, 1,523 at step 20, 91 at step 30, 0 at step 40.
Visual motion between simulated states is interpolated by the renderer. Video appearance is illustrative and does not constitute a measurement of independent individual movement.
Attribution
Qwen3.5-4B: Qwen, Apache-2.0.
QwenGenesis ecosystem, experiments and demos: @suayptalha