Antonio Gemma3 Evo Q4 is not just another quantized LLM. It's a self-learning micro-intelligence with EvoMemory™, RAG-Lite, and auto-evolution capabilities, optimized for Raspberry Pi 4 & 5 and tested for production 24/7 deployment.

Version: v0.5.0 (NEW: Adaptive Prompting) Author: Antonio Consales (antconsales) Base Model: Google Gemma 3 1B IT

💻 Production-Ready for Raspberry Pi

✅ Tested on Raspberry Pi 4 (4GB) — 3.32 t/s sustained (100% reliable over 60 minutes) ✅ Fully offline — no external APIs, no internet required ✅ Self-learning — EvoMemory™ saves neurons from every conversation ✅ Bilingual — seamlessly switches between Italian and English ✅ 24/7 deployment tested — Zero failures in 60-minute soak test

🔍 What Makes It Special

Unlike traditional quantized models, Antonio Gemma3 Evo Q4 learns and evolves:

🧬 EvoMemory™ — Saves "neurons" with input, output, confidence, and mood
🔍 RAG-Lite — Retrieves past experiences using BM25 (no FAISS!)
🎯 Self-evaluation — Assigns confidence scores (0-1) to every response
🌱 Auto-evolution — Generates new reasoning rules from accumulated neurons
🔒 100% Offline — Runs completely local on Raspberry Pi 4 (4GB RAM)
🌐 Bilingual — Auto-detects IT/EN and responds in the same language
⚡ Fast — 3.32 tokens/s sustained on Pi 4 with Q4_K_M quantization
🎯 NEW: Adaptive Prompting — Smart question classification (SIMPLE/COMPLEX/CODE/CREATIVE) for 3.6x speedup on simple queries

"The little brain that grows with you" 🧠

📊 Benchmark Results (Updated Oct 21, 2025)

Complete 60-minute soak test on Raspberry Pi 4 (4GB RAM) with Ollama.

Production Metrics

Metric	Value	Status
Sustained throughput	3.32 t/s (256 tokens)	✅ Production-ready
Reliability	100% (455/455 requests)	✅ Perfect
Avg response time	7.92s	✅ Consistent
Thermal stability	70.2°C avg (max 73.5°C)	✅ No throttling
Memory usage	42% (1.6 GB)	✅ No leaks
Uptime tested	60+ minutes continuous	✅ 24/7 ready

Performance by Token Count

Tokens	Run 1	Run 2	Run 3	Average*
128	0.24	3.44	3.45	3.45 t/s
256	3.43	3.09	3.43	3.32 t/s
512	2.76	2.77	2.11	2.55 t/s

*Average excludes cold-start (first run)

Model Comparison

Model	Size	Sustained Speed	Recommended For
Q4_K_M ⭐	769 MB	3.32 t/s	Production (tested 60min, 100% reliable)
Q4_0	687 MB	3.45 t/s	Development (faster but less stable)

📊 View Complete Benchmark Report — Full performance, reliability, and stability analysis

Benchmark Methodology

Duration: 60.1 minutes (3,603 seconds)
Total requests: 455 (2-second interval)
Platform: Raspberry Pi 4 (4GB RAM, ARM Cortex-A72 @ 1.5GHz)
Runtime: Ollama 0.3.x + llama.cpp backend
Monitoring: CPU temp, RAM usage, load average (sampled every 5s)
Tasks: Performance (128/256/512 tokens), Quality (HellaSwag, ARC, TruthfulQA), Robustness (soak test)

Recommendation: Use Q4_K_M for production deployments (proven 100% reliability over 60 minutes). Use Q4_0 for development/testing if you need slightly faster inference.

🧩 Available Models

This repository contains two quantization variants:

gemma3-1b-q4_0.gguf (≈687 MB) — Faster, 3% higher throughput, suitable for development
gemma3-1b-q4_k_m.gguf (≈769 MB) — Better quality, production-tested for 60+ minutes

🎯 Important: Two Usage Modes

Mode 1: Ollama Only (Simple Inference) ⚡

Download the GGUF model and run with Ollama:

bash

ollama pull antconsales/antonio-gemma3-evo-q4
ollama run antconsales/antonio-gemma3-evo-q4

What you get:

✅ Fast inference (3.32 t/s on Pi 4)
✅ Bilingual chat (IT/EN)
✅ Offline, privacy-first
❌ NO EvoMemory (doesn't save conversations)
❌ NO RAG (doesn't retrieve past experiences)
❌ NO auto-evolution (doesn't generate rules)

Best for: Quick tests, one-off questions, simple chatbot

Mode 2: Full Evolution Stack (Self-Learning) 🧠

For EvoMemory™, RAG-Lite, and auto-evolution, use the full Python stack from GitHub:

bash

git clone https://github.com/antconsales/antonio-gemma3-evo-q4.git
cd antonio-gemma3-evo-q4
bash scripts/install.sh
uvicorn api.server:app --host 0.0.0.0 --port 8000

What you get:

✅ EvoMemory™ — Saves neurons from every conversation
✅ RAG-Lite — Retrieves past experiences (BM25)
✅ Auto-evolution — Generates reasoning rules over time
✅ Confidence scoring — Knows when it's uncertain
✅ FastAPI server — REST + WebSocket endpoints

Comparison:

Feature	Ollama Only	Full Stack
Inference speed	3.32 t/s	3.32 t/s
Learns from chats	❌	✅ EvoMemory
Retrieves memories	❌	✅ RAG-Lite
Generates rules	❌	✅ Auto-evolution
API endpoints	❌	✅ FastAPI
Setup time	1 min	5 min

🛠️ Quick Start Options

Option 1: Ollama Only (see Mode 1 above)

Option 2: Load Directly from GGUF

bash

# Download model from HuggingFace
wget https://huggingface.co/chill123/antonio-gemma3-evo-q4/resolve/main/gemma3-1b-q4_k_m.gguf

# Create Modelfile
cat > Modelfile <<'EOF'
FROM ./gemma3-1b-q4_k_m.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 1024
PARAMETER num_thread 4
PARAMETER repeat_penalty 1.05
PARAMETER stop "<end_of_turn>"
PARAMETER stop "</s>"

SYSTEM """You are Antonio, an offline AI assistant running on a Raspberry Pi. You MUST detect the user's language and respond in the SAME language:

- If the user writes in Italian, respond ONLY in Italian
- If the user writes in English, respond ONLY in English

You are helpful, friendly, and concise. When you're uncertain, you admit it instead of guessing."""
EOF

# Create and run model
ollama create antonio-evo -f Modelfile
ollama run antonio-evo

🚀 Quick Start with Full Evolution Stack

For the complete self-learning system with EvoMemory™, RAG-Lite, and auto-evolution:

bash

# Clone the full project
git clone https://github.com/antconsales/antonio-gemma3-evo-q4.git
cd antonio-gemma3-evo-q4

# Install and run
bash scripts/install.sh
python -m api.server

Visit: http://localhost:8000/docs for interactive API documentation

Features of the full stack:

EvoMemory™ SQLite database
RAG-Lite with BM25 search
Confidence auto-evaluation
Rule regeneration (auto-evolution)
FastAPI server with WebSocket support
MCP-compatible tool system

💡 Key Features

1️⃣ EvoMemory™ — Living Memory

Every conversation creates a neuron:

json

{
  "id": 123,
  "input_text": "Accendi il LED rosso",
  "output_text": "OK, attivo GPIO 17 su HIGH",
  "confidence": 0.85,
  "mood": "positive",
  "user_feedback": 1,
  "skill_id": "gpio_control",
  "timestamp": "2025-10-21T14:30:00Z"
}

Features:

Auto-pruning of low-confidence old neurons
Neuron compression (similar patterns → meta-neurons)
Context-aware retrieval via hash matching

2️⃣ RAG-Lite — Pure Python BM25

No FAISS, no ChromaDB. Just:

BM25 scoring (Okapi formula)
SQLite full-text search (FTS5)
Top-K retrieval with confidence threshold
Zero external dependencies

3️⃣ Auto-Evolution — Rule Regeneration

Every N conversations (configurable), Antonio:

Analyzes high-confidence neurons
Extracts reasoning patterns
Generates new rules (e.g., "If user asks time-sensitive question → check recency")
Saves rules to instinct.json

🏗️ Architecture

┌─────────────────────────────────────────────┐
│  Antonio Gemma3 Evo Q4 - Evolution Layer    │
├─────────────────────────────────────────────┤
│                                             │
│  ┌──────────────┐      ┌─────────────────┐ │
│  │ EvoMemory™   │◄────►│  RAG-Lite       │ │
│  │ (SQLite)     │      │  BM25 Search    │ │
│  └──────────────┘      └─────────────────┘ │
│         ▲                      ▲            │
│         │                      │            │
│  ┌──────┴──────────────────────┴──────────┐ │
│  │    Inference Engine (llama.cpp)        │ │
│  │    • Q4_0 / Q4_K_M                     │ │
│  │    • Optimized for Pi 4                │ │
│  └────────────────────────────────────────┘ │
│         ▲                      │            │
│         │                      ▼            │
│  ┌──────┴─────────┐    ┌──────────────────┐ │
│  │ Action Broker  │    │  Confidence      │ │
│  │ (MCP-ready)    │    │  Auto-Eval       │ │
│  └────────────────┘    └──────────────────┘ │
│                                             │
│  FastAPI Server (REST + WebSocket)         │
└─────────────────────────────────────────────┘

🎯 Use Cases

Recommended for:

✅ Home AI assistants (24/7 operation)
✅ IoT edge inference (low power budget)
✅ Offline chatbots (privacy-first)
✅ Educational projects (affordable hardware)
✅ Voice assistants (bilingual IT/EN)
✅ Self-learning experiments (neuron/rule evolution)

Not recommended for:

❌ Real-time applications (<500ms latency)
❌ Batch processing (CPU-bound, single-threaded)
❌ High concurrency (>5 simultaneous users)

📚 Links

GitHub (Full Stack): https://github.com/antconsales/antonio-gemma3-evo-q4
Ollama: https://ollama.com/antconsales/antonio-gemma3-evo-q4
HuggingFace: https://huggingface.co/chill123/antonio-gemma3-evo-q4
Donate: https://www.paypal.com/donate/?business=58ML44FNPK66Y&currency_code=EUR

📄 License

This model is licensed under the Gemma License Agreement (inherits from base model).

The evolution stack code (EvoMemory™, RAG-Lite, etc.) is dual-licensed:

Gemma License (model weights)
MIT License (Python code)

See LICENSE for details.

Built with ❤️ for offline AI and edge computing

"Il piccolo cervello che cresce insieme a te" — Antonio Gemma3 Evo

Support ethical, local, and independent AI. Every donation helps Antonio Gemma grow and evolve. 💙