OPENFLUKE • AI INFRASTRUCTURE LAB
Sovereign Intelligence
for the Edge.
Own your intelligence. Build, deploy, and control AI entirely on your own infrastructure.
One Engine. Two Paths.
"Primecraft for Development. Weave for Business Automation."
109 Tasks Solved on ARC-AGI-1
Ensemble Fusion v2
Loom's Power Features
-
🔲
Grid Architecture Data flows through a 2D grid of cells
-
🧠
Parallel Brains MHA, LSTM, RNN, CNN running together
-
🌐
13 Platform Builds One model, copy-paste anywhere — no conversion needed
🚀 v0.0.8: Windows/Mac/Linux (x86_64 + ARM64) • Android ARM64 • iOS ARM64 • WASM
🔮 In Development: WebGPU acceleration — GPU power on any hardware 📦 See the Release -
🔍
Live Debugging Watch every neuron fire in real-time
What is Loom?
Loom is a Deterministic Neural Virtual Machine (DNVM). It serves as a portable execution environment that guarantees bitwise-identical results across all platforms—whether you're running on a Linux server, a Windows desktop, or inside a web browser.
JSON network configs act as "bytecode". Define your neural architecture once, and execute it anywhere without code changes.
Runtime generation of WebGPU shaders allows for high-performance compute without heavy external dependencies.
Achieve 0.0000000000 difference between CPU and GPU, x86 and ARM. Loom enforces determinism by design.
Sine Wave Adaptation
Testing continuous adaptation as the target function shifts frequency every 2.5 seconds.
📄 View Source Code| Mode | Accuracy | Stability | Consistency | Throughput | Score |
|---|---|---|---|---|---|
| NormalBP | 91.3% | 87.2% | 98.5% | 75,568 | 6,493 |
| StepBP | 79.4% | 86.6% | 95.5% | 23,611 | 1,954 |
| Tween | 47.4% | 82.4% | 51.0% | 170,699 | 7,170 |
| TweenChain | 60.4% | 84.5% | 87.5% | 171,861 | 12,709 |
| StepTween | 54.3% | 74.8% | 64.5% | 313,465 | 15,120 |
| StepTweenChain 🏆 | 55.4% | 84.9% | 71.0% | 312,736 | 18,841 |
Six Training Modes
Choose based on your use case. See benchmark comparison above.
NormalBP Score: 6,493
Batch backprop. PAUSES to train. Maximum accuracy.
- Model pre-training
- Benchmark competitions
- Offline dataset training
StepBP Score: 1,954
Per-sample gradients. Online learning.
- Streaming data
- Reinforcement learning
- Low-latency updates
Tween Score: 7,170
Weight interpolation toward targets.
- Model fine-tuning
- Transfer learning
- Smooth weight merging
TweenChain Score: 12,709
Layered tweening with chain rule.
- Deep network adaptation
- Layer-wise fine-tuning
- Gradient-aware merging
StepTween Score: 15,120
Hybrid step + tween. Max throughput.
- High-frequency trading
- Video game AI
- Sensor stream processing
StepTweenChain 🏆 Score: 18,841
Full hybrid. Best combined score.
- Real-time robotics
- Adaptive control systems
- Dynamic environment agents
Primecraft: The AI Gym
We built a game universe just to prove our AI works. Spawn an agent. Teach it to follow you. Then change the gravity and watch it rewrite its own brain in real-time.
Alpha v0.30.0 — Model Sharing Patch
Runtime Gallery
