https://github.com/larionovavi-stack/genesis2-cascade-moe

Genesis 2 — Cascade MoE Neural Network | Patented CPU-only AI: 10,800 experts, 100% accuracy, 18ms inference, zero forgetting | No GPU required | Self-hosted alternative to GPT/LLaMA | Network automation & DevOps | Python/PyTorch
https://github.com/larionovavi-stack/genesis2-cascade-moe

ai artificial-intelligence bilingual cascade-moe cpu-only cybersecurity deep-learning devops linux-administration machine-learning mixture-of-experts network-automation neural-network nlp no-gpu on-premise patented pytorch self-hosted sysadmin

Last synced: 3 days ago
JSON representation

Host: GitHub
URL: https://github.com/larionovavi-stack/genesis2-cascade-moe
Owner: larionovavi-stack
License: other
Created: 2026-06-13T11:00:22.000Z (10 days ago)
Default Branch: main
Last Pushed: 2026-06-13T13:08:42.000Z (10 days ago)
Last Synced: 2026-06-13T13:09:33.133Z (10 days ago)
Topics: ai, artificial-intelligence, bilingual, cascade-moe, cpu-only, cybersecurity, deep-learning, devops, linux-administration, machine-learning, mixture-of-experts, network-automation, neural-network, nlp, no-gpu, on-premise, patented, pytorch, self-hosted, sysadmin
Language: HTML
Homepage: https://avlarion.gumroad.com
Size: 142 KB
Stars: 0
Watchers: 0
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md
- License: LICENSE

Awesome Lists containing this project

awesome-iot - Genesis 2 - Cascade MoE neural network for IoT edge deployments. CPU-only inference (18ms), no GPU required, patented architecture with zero catastrophic forgetting. (AI / Tutorial)
Awesome-Efficient-Inference-for-LRMs - link

README

          


  



Genesis 2 — Cascade MoE Neural Network




  The World's First Patented Neural Architecture That Runs on CPU





  

  

  

  

  

  





  Academic $299 •

  Professional $1,499 •

  Enterprise $4,999 •

  Source + Patent Bundle $5,000 •

  Interactive Reference Guide





  



> **[Try the Live Demo](https://gist.github.com/larionovavi-stack/d9bdc484813df7cb488a842cb4a0cd62)** — click the link in the Gist for the current demo URL. Model hosted on [Kaggle](https://www.kaggle.com/datasets/alexanderlar/genesis2-cascade-moe-model). If the demo is unavailable, email **avlarionov@hotmail.com** to request a restart.

---

## What is Genesis 2?

Genesis 2 is a **fundamentally new neural network architecture** that eliminates the need for GPU, external LLMs, and massive compute resources. It uses Cascade Activation of a Shared Neuron Pool — a patented approach where experts share neurons instead of duplicating parameters.

**No GPU. No Cloud. No API costs. No token limits. Runs on your laptop.**

```

Traditional MoE:  Expert₁[500MB] + Expert₂[500MB] + ... = 50GB+, GPU required

Genesis 2:        Expert₁[route] + Expert₂[route] + ... = 3.64 GB total, CPU only

                  ↑ shared neuron pool, each expert is just a list of neuron IDs

```

## Why Genesis 2?

| Traditional AI (GPT, LLaMA, etc.) | Genesis 2 |

|:---|:---|

| $2,000+/mo GPU costs | **$0** — runs on CPU |

| API rate limits & downtime | **Unlimited** — self-hosted |

| Data leaves your network | **100% on-premise** |

| Catastrophic forgetting | **Zero forgetting** — mathematically guaranteed |

| Minutes to fine-tune | **130ms** to learn a new fact |

| Token window limits (4K-128K) | **Infinite context** — no limits |

| Vendor lock-in | **You own the code** |

## Quick Start

```bash

# Install dependencies

pip install torch numpy requests

# Start the web server

python genesis2_web.py

# Open in browser

open http://localhost:8765

```

## API

```python

import requests

API = "http://localhost:8765"

# Ask a question (returns answer + executable commands)

r = requests.post(f"{API}/api/query", json={"question": "configure nginx reverse proxy"})

print(r.json()["answer"])

print(r.json()["commands"])

# Teach new knowledge (learns in 130-550ms)

requests.post(f"{API}/api/learn", json={

    "question": "how to restart Apache",

    "answer": "Restart Apache web server",

    "exec": "systemctl restart apache2"

})

# Save state

requests.post(f"{API}/api/save")

```

## What's New in v1.1 

Released: **June 2026**

| Feature | Description |

|:--------|:------------|

| 🧠 **Neuron Splitting** (Patent п.5) | Overloaded neurons auto-split via 2-means clustering. Coherence threshold 0.40 triggers split → two child neurons inherit parent weights |

| 💬 **Dialogue Context** | Model tracks conversation state: "no thanks", "nothing needed", "пока ничего" → correct conversational replies instead of technical routing |

| 🔧 **Command Substitution** | Auto-fills IP/port/subnet from user's question into exec commands: `ping 10.0.0.1` → `ping -c 4 10.0.0.1` |

| 🔤 **Typo Normalization** | Repeated Cyrillic letters collapsed: "ппривет" → "привет", "приввет" → "привет" (Latin preserved: "need" stays "need") |

| 📊 **111/111 Test Suite** | Extended benchmark from 30 to **111 queries** across 43 topics: networking, security, Docker, Cisco, VPN, DNS, databases, monitoring, SCADA, VoIP and more |

| 🌐 **Bilingual 100%** | Both RU and EN at 100% accuracy simultaneously — verified across all 43 topic categories |

## Benchmarks

| Metric | v1.0 | **v1.1** |

|:-------|:-----|:---------|

| Shared Neurons | 12,100+ | **12,651** |

| Trained Experts | 10,800+ | **10,800+** |

| Test accuracy | 100% (30/30) | **100% (111/111)** |

| Topics covered | 15 | **43** |

| Inference latency | 18-27ms | **18-27ms** |

| Learning speed | 130-550ms | **130ms** per fact |

| Zero forgetting (cosine) | 1.000000 | **1.000000** |

| Neuron splitting | ✗ | **✓ (auto)** |

| Dialogue context | ✗ | **✓** |

| Command substitution | ✗ | **✓** |

| RAM usage | 3.5GB | **3.64 GB** |

| GPU required | No | **No** |

### Test Results v1.1 — 111/111 across 43 topics

```

networking RU/EN  ✅✅✅✅✅✅✅✅✅✅✅  (11/11)

linux RU/EN       ✅✅✅✅✅✅✅✅✅✅✅  (11/11)

security RU/EN    ✅✅✅✅✅✅✅✅✅  (9/9)

vpn RU/EN         ✅✅✅✅✅  (5/5)

docker/k8s RU/EN  ✅✅✅✅✅✅✅✅  (8/8)

cisco RU/EN       ✅✅✅✅✅  (5/5)

dns/dhcp RU/EN    ✅✅✅✅✅✅  (6/6)

monitoring RU/EN  ✅✅✅✅✅  (5/5)

databases         ✅✅✅✅  (4/4)

nginx/web         ✅✅✅✅✅  (5/5)

windows           ✅✅  (2/2)

mikrotik          ✅✅  (2/2)

voip/sip          ✅✅  (2/2)

scada/iot         ✅✅  (2/2)

backup            ✅✅  (2/2)

devops            ✅✅✅✅  (4/4)

troubleshooting   ✅✅✅✅  (4/4)

cloud/virt        ✅✅✅  (3/3)

macos             ✅✅✅  (3/3)

traffic           ✅✅✅  (3/3)

greetings/typos   ✅✅✅✅✅✅✅  (7/7)

slang/infra       ✅✅✅✅  (4/4)

                           ───────

TOTAL:            ✅ 111/111 = 100%

```

## Architecture

Genesis 2 is built on 8 patented innovations:

### 1. Shared Neuron Pool

All neurons live in a single shared pool. Experts don't have their own parameters — they reference neurons by ID. One neuron can serve 50+ experts simultaneously. This makes the model **100x smaller** than traditional MoE.

### 2. Expert as Route

Each expert is just a list of neuron IDs — a "route" through the shared pool. Adding a new expert costs **bytes, not megabytes**. 10,800+ expert routes fit in 3.64 GB.

### 3. Cascade Activation (No Router)

Traditional MoE uses a trained router to pick experts. Genesis 2 uses a reverse index (neuron → experts) to find relevant experts in **0.14ms**. No router training, no routing errors.

### 4. One-Step Learning

To learn a new fact: freeze all shared neurons, create a new expert with a micro-head. Takes **130-550ms**. The new knowledge never interferes with existing knowledge.

### 5. Zero Catastrophic Forgetting

Each expert has its own micro-head (output layer). New experts can't modify existing ones. **Mathematically guaranteed** — cosine similarity = 1.000000 before/after learning.

### 6. Hash Neuron Embedding

Custom embedding system with 9,761 tokens across 72 types. No dependency on external models (MiniLM, BERT, etc.). Fully self-contained.

### 7. Infinite Context

Every learned fact becomes a permanent expert. No token window limits. 10,000 facts = 10,000 experts, all accessible instantly.

### 8. Native Generation via Concept Chains

Output is generated through a composer that chains related concepts from activated experts. Not template matching — actual generation.

```

Input → Hash Embedding (512d) → ANN Search → Seed Experts

     → Cascade Activation → Shared Neuron Pool → Composer → Output

```

## Knowledge Domains (35)

> The model is fully bilingual (RU + EN). Trained on 35 domains with 100% accuracy in both languages. Genesis 2 learns new facts in **130ms** — you can train your own model on any language and any domain in minutes, not days.

Networking (Cisco, MikroTik)Linux AdministrationDocker & Kubernetes

Security & HardeningWiFi ConfigurationDNS/DHCP/BIND

VPN (WireGuard, OpenVPN)Databases (PostgreSQL, MySQL)Web Servers (Nginx, Apache)

Monitoring (Zabbix, Prometheus)DevOps (Ansible, Terraform)Python Scripting

Bash AutomationPacket AnalysisVoIP (Asterisk)

Windows Active DirectorymacOS AdministrationVirtualization

SCADA/ICSCloud (AWS/GCP/Azure)Server Configuration

Mobile Protocols

## System Requirements

| Component | Minimum | Recommended |

|:----------|:--------|:------------|

| CPU | Any modern (ARM or x86) | 4+ cores |

| RAM | 6 GB | 16 GB |

| Disk | 4 GB | 10 GB |

| Python | 3.9+ | 3.11+ |

| PyTorch | 2.0+ | 2.3+ |

| OS | macOS / Linux / Windows | Any |

| GPU | **Not required** | Not required |

## Patent

**Status:** Filed at FIPS Russia, 31.05.2026

**Type:** Utility Model, IPC G06N 3/04

**Claims:** 2 independent + 6 dependent (8 total)

**RCIS Blockchain Certificate:** #1823-376-572

The Cascade MoE architecture is protected by a pending patent. The patent covers all 8 architectural innovations listed above.

## OS-Aware Execution

Genesis 2 detects the host operating system and adapts:

- **macOS**: Strips `sudo`, warns about Linux-only commands, uses macOS equivalents

- **Linux**: Full command execution with `sudo` support

- **Windows**: Suggests PowerShell alternatives

- **Safety**: Blocks dangerous commands (`rm -rf`, `mkfs`, `dd`, `shutdown`)

## Editions

| Edition | Price | License | Includes |

|:--------|:------|:--------|:---------|

| [**Academic**](https://avlarion.gumroad.com/l/lqtsbo) | $299 | 1 person, research only | Source + model + docs |

| [**Professional**](https://avlarion.gumroad.com/l/vrzudu) | $1,499 | 5 users, commercial | + 30 datasets + 12mo updates |

| [**Enterprise**](https://avlarion.gumroad.com/l/atmon) | $4,999 | Unlimited, commercial | + patent docs + book + lifetime updates |

| [**Source + Patent Bundle**](https://avlarion.gumroad.com/l/ymyagw) | $5,000 | White-label rights | + patent license + 5h consultation |

## Project Structure

```

genesis2-cascade-moe/

├── genesis2_core.py          # Core: neurons, cascade, shared pool, training

├── genesis2_gen.py           # Generation: concept chains, composer, boost

├── genesis2_agent.py         # Agent: learn/reason/plan/chat/self-learn

├── genesis2_web.py           # Web UI + REST API + OS detection

├── genesis2_repl.py          # Interactive terminal REPL

├── embedding/

│   └── train_embedding.py    # Custom hash embedding training

├── datasets/                 # 30 training datasets (Professional+)

├── PATENT/                   # Patent materials (Enterprise+)

└── requirements.txt

```

## Author

**Larionov Alexander Viktorovich** (Ларионов Александр Викторович)

- SCADA/ICS Engineer with 10+ years of industrial automation experience

- AI Researcher specializing in novel neural architectures

- Patent holder (Cascade MoE, FIPS Russia 2026)

**Contact:** avlarionov@hotmail.com

**GitHub:** [larionovavi-stack](https://github.com/larionovavi-stack)

**Products:** [avlarion.gumroad.com](https://avlarion.gumroad.com)

## Also by Author

- **[atwSCADA](https://github.com/larionovavi-stack/awtscada)** — Free SCADA system in a single HTML file (IEC 61850, OPC UA, Modbus TCP)

- **[Network Automation with AI](https://github.com/larionovavi-stack/network-automation-ai-guide)** — 132-page practical guide ($29)

## Affiliate Program

Earn **40% commission** on every sale by promoting Genesis 2.

**[→ Join the Affiliate Program](https://avlarion.gumroad.com/affiliates)**

Payouts via Gumroad. No approval required — instant access.

## License

This repository contains the documentation, architecture description, and demo materials. The full source code and trained model are available through [Gumroad](https://avlarion.gumroad.com).

Patent pending. All rights reserved. (c) 2026 Larionov Alexander Viktorovich.

---



  No GPU. No Cloud. No Limits.


  Get Genesis 2 Academic — $299

ecosyste.ms

Data

Tools

Indexes

Applications

Experiments

Awesome

https://github.com/larionovavi-stack/genesis2-cascade-moe

Awesome Lists containing this project

README

Genesis 2 — Cascade MoE Neural Network