---
language:
- en
library_name: peft
license: apache-2.0
tags:
- lora
- conversational
- text-generation
- peft
- fine-tuned
- mamba
- ssm
- neuralai
- base-model
model_id: Subject-Emu-5259/NeuralAI
inference: false
---
# π§ NeuralAI: The Generative AI Engine
---
## π Repository Composition
| Language | Percentage |
| --- | --- |
| Python | 71.1% |
| HTML | 13.0% |
| JavaScript | 12.4% |
| CSS | 2.6% |
| Shell | 0.4% |
| Jupyter Notebook | 0.3% |
| Jinja | 0.2% |
**The High-Velocity AI Engine for Your Entire Vibe Stack**
NeuralAI is the central intelligence engine developed by **De'Andrew Preston Harris**. Conceived and engineered as an owned AI platform, it spans fine-tuned transformer models, custom SSM base models, DPO alignment, and a production web UI β all designed for local-first, private AI computing.
---
## ποΈ Model Family
```mermaid
graph TB
subgraph "NeuralAI Model Family"
direction TB
K1["𧬠Mamba K1
130M Β· SFT 50 steps
First Owned Base"]
K2["𧬠Mamba K2
793M Β· Q4_K_M GGUF
460MB Β· Production"]
K3["π¬ Mamba K3
SFT 500-1000 steps
10K+ UltraChat
In Training"]
K1 --> K2 --> K3
K2 --> PROD["π Active Inference
LM Studio Β· 460MB RAM
neuralai-web-ui.zocomputer.io"]
end
style K1 fill:#4a90d9,color:#fff
style K2 fill:#22c55e,color:#fff
style K3 fill:#f59e0b,color:#000
style PROD fill:#10b981,color:#fff
```
### Complete Lineup
| Model | Architecture | Params | Training | Status | Location |
|-------|-------------|--------|----------|--------|----------|
| **Mamba K1** | Mamba SSM | 130M | SFT LoRA 500 steps on 1K UltraChat (intel format) | π Retraining for chat coherence | `Subject-Emu-5259/NeuralAI-Mamba-K1` |
| **Mamba K2** | Mamba SSM | 793M | Base pretrained β SFT queued (Q4_K_M GGUF) | β οΈ Base model only | `Subject-Emu-5259/NeuralAI-Mamba-K2` |
| **Mamba K3** | Mamba SSM | 2.8B | Base pretrained β SFT queued | β οΈ Base model only | local `models/mamba-k3-base/` |
### Why Mamba SSM
| Property | Benefit |
|----------|--------|
| Complexity | \(O(n)\) linear β scales to long context efficiently |
| Inference | Fast at any sequence length, not just short prompts |
| Memory | K2 runs at 460MB (Q4_K_M GGUF) β fits on any device |
| Ownership | NeuralAI trains and merges all release weights on top of public Mamba SSM bases β every released GGUF is a fully merged model, not a raw base |
| Ecosystem | LM Studio, llama.cpp, Hugging Face β mature deployment options |
---
## π Vision & Manifesto
NeuralAI doesn't just predict text; it *operates the work*. The core mission is to create a multimodal generative system that bridges the gap between raw idea and execution. By fusing autoregressive generation with adaptive agency, NeuralAI becomes more than a chatbot β it is a persistent, reasoning partner.
Born from resilience and ambition in Memphis, Tennessee and West Memphis, Arkansas, NeuralAI represents a forward-thinking approach to personal, private AI computing.
---
## π οΈ Tech Stack & Architecture
NeuralAI is built on a high-performance architecture that decouples the inference engine from the web interface, enabling lightweight cloud hosting with powerful local inference.
### Core Stack
- **Production Model**: Mamba K2 793M Q4_K_M GGUF (460MB) via llama.cpp β active inference engine
- **Model Family**: Mamba K1 (130M, SFT LoRA retry) β Mamba K2 (793M base GGUF) β Mamba K3 (2.8B base)
- **Inference Engine**: llama.cpp server with a custom `neuralai-intel` chat format (vocabulary-friendly for GPT-NeoX / Mamba tokenizers)
- **Vocal Identity**: Andrew (Warm/Multilingual) β Optional voice synthesis integration
- **Web Interface**: Custom Flask UI served via Zo Computer at `neuralai-web-ui-deandrewharris.zocomputer.io`
- **Tool Chain**: 10 live slash commands (/web, /fetch, /browse, /research, /img, /speak, /summarize, /translate, /news, /yt) + NLβTool Router
### Future Scale Path
| Stage | Params | Goal | Status |
|-------|--------|------|--------|
| **Mamba K1 SFT v2** | 130M | Finish 500-step SFT with intel format + GGUF | π Active |
| **Mamba K2 SFT** | 793M | SFT LoRA 500 steps β merge β Q4_K_M GGUF | π Next |
| **Mamba K3 SFT** | 2.8B | SFT LoRA 1000 steps β merge β Q4_K_M GGUF | π Next |
| **Mamba 2B/3B** | ~2-3B | Scaled SSM architecture, benchmarks | π Planned |
---
## β¨ Key Features & Capabilities
### π¬ Multimodal Chat & Agentic Intelligence
- **High-Velocity Text Inference**: Fast, local inference with deep context awareness
- **Deep Reasoning Mode**: Integration of test-time compute and chain-of-thought reasoning
- **Autonomous Agentic Workflows**: Agent-mode interaction with browser, terminal, and third-party apps
- **Live S2S (Speech-to-Speech)**: Real-time voice interaction with integrated microphone interface
- **Identity Vault & Memory**: Persistent user memory and rule constraints
### π§ Developer & Engineering Tools
- **10 Web Tool Commands**: Search, fetch, browse, research, image gen, TTS, summarize, translate, news, YouTube
- **NLβTool Router**: Natural language web requests auto-routed to the right tool
- **Model Manager**: CLI switching between all registered models
- **Benchmark Suite**: Perplexity, generation diversity, MMLU-style, reasoning tests
---
## π Model Lineage
```mermaid
timeline
title NeuralAI Model Evolution
: Custom 135M base SFT
2026 Q3 : Mamba K1 β First owned base
: 130M SSM Β· Proof of Concept
2026 Q3 : Mamba K2 β Scaled base
: 790M Q4_K_M Β· GGUF ready
2026 Q3 : Mamba K3 β Full SFT
: 500-1000 steps Β· 10K+ samples
2026 Q4 : Mamba 2B/3B β Next scale targets
: 2B SSM β 3B Core Intelligence
```
---
## π Deployment
```bash
# 1. Start the inference service
cd NeuralAI
supervisorctl -c /etc/zo/supervisord-user.conf restart neuralai-lmstudio
# 2. Start the web UI service
python3 services/webui_service.py
```
### Mamba K2 (LM Studio / llama.cpp)
```bash
# Download from HuggingFace
huggingface-cli download Subject-Emu-5259/NeuralAI-Mamba-K2 \
mamba-790m-hf.Q4_K_M.gguf --local-dir ./models/
# Serve with the NeuralAI chat format
python3 services/lmstudio_server.py \
--model models/mamba-790m-hf.Q4_K_M.gguf \
--chat_format neuralai-intel \
--port 1234
```
### Mamba K1 (Python)
```python
from transformers import MambaForCausalLM, AutoTokenizer
model = MambaForCausalLM.from_pretrained("Subject-Emu-5259/NeuralAI-Mamba-K1")
tokenizer = AutoTokenizer.from_pretrained("Subject-Emu-5259/NeuralAI-Mamba-K1")
```
### Containerized Deployments
| Deployment | Dockerfile | Stack | Status |
| --- | --- | --- | --- |
| **Web Chat** | `services/start_lmstudio.sh` + Flask UI | llama.cpp + `neuralai-web-ui.zocomputer.io` | β
Live |
---
## π NeuralAI Ecosystem
The standalone software implementation of the NeuralAI core is **NeuralLabs**:
π [github.com/Subject-Emu-5259/NeuralLabs](https://github.com/Subject-Emu-5259/NeuralLabs)
**Software Downloads**: Latest beta builds available at:
π [zo.pub/deandrewharris/neurallabs-beta](https://zo.pub/deandrewharris/neurallabs-beta)
---
## π Current State & Active Goals
- **Legacy DPO v17 / Air 135M / SmolLM2-360M**: Retired and removed from the repository
- **Mamba K1**: First owned SSM base model β retraining with intel-format SFT for coherent chat (Colab/GPU)
- **Mamba K2**: Base pretrained GGUF ready β awaiting SFT
- **Mamba K3**: 2.8B base downloaded β awaiting SFT
- **Last Maintenance**: August 1, 2026 (Mamba Era β fix chat format + retrain pipeline)
---
## π€ Creator
Built by **De'Andrew Preston Harris** ([@deandrewharris94](https://linkedin.com/in/deandrewharris94/)) with Google Gemini AI Studio/Colab collaboration.
From Memphis, Tennessee. Raised in West Memphis, Arkansas. AI Software Engineering at Maestro College.
---
# NeuralAI β Hugging Face sync is live