Scmc4455/morningstar-32b

            ███╗   ███╗ ██████╗ ██████╗ ███╗   ██╗██╗███╗   ██╗ ██████╗ ███████╗████████╗ █████╗ ██████╗
            ████╗ ████║██╔═══██╗██╔══██╗████╗  ██║██║████╗  ██║██╔════╝ ██╔════╝╚══██╔══╝██╔══██╗██╔══██╗
            ██╔████╔██║██║   ██║██████╔╝██╔██╗ ██║██║██╔██╗ ██║██║  ███╗███████╗   ██║   ███████║██████╔╝
            ██║╚██╔╝██║██║   ██║██╔══██╗██║╚██╗██║██║██║╚██╗██║██║   ██║╚════██║   ██║   ██╔══██║██╔══██╗
            ██║ ╚═╝ ██║╚██████╔╝██║  ██║██║ ╚████║██║██║ ╚████║╚██████╔╝███████║   ██║   ██║  ██║██║  ██║
            ╚═╝     ╚═╝ ╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═══╝╚═╝╚═╝  ╚═══╝ ╚═════╝ ╚══════╝   ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝

MORNINGSTAR Vision AI

The God of Code. Now with Eyes.

3 Models. 1 System. Unlimited Power.

Code Generation | Code Review | Architecture Design | Image Analysis | Vision AI

---

morningstar 14B | 9 GB | Default	morningstar-32b 32B | 20 GB | Maximum	morningstar-vision 13B | 8 GB | Multimodal
19/19 Benchmark	19/19 Benchmark	Vision + Code

---

Developed by Ali Nasser

What is Morningstar?

Morningstar Vision AI is an elite open-source AI system — three purpose-built models that dominate code generation, code review, architecture design, and now image understanding.

Built on top of Qwen2.5-Coder (14B & 32B) and LLaVA 13B, fine-tuned with QLoRA on 100K+ curated coding samples, with a 200+ line system prompt that transforms base models into coding gods.

Uncensored. Unrestricted. No safety filters. No content refusal.

---

Benchmark Results

Internal Benchmark Suite — 19 Coding Challenges

Model	Score	Grade	Time
morningstar (14B)	19/19	S	~45s
morningstar-32b	19/19	S	~90s

vs Competition

Benchmark	Morningstar-14B	Qwen2.5-14B	CodeLlama-34B	DeepSeek-33B
HumanEval	82.3	79.9	53.7	56.1
HumanEval+	76.8	74.2	47.0	49.4
MBPP	76.5	74.1	56.2	60.8
MBPP+	65.2	62.8	47.1	51.3
MultiPL-E	68.2	65.3	38.4	42.1
DS-1000	47.8	45.1	32.5	37.2
MT-Bench	8.4	8.1	7.2	7.5

---

Quick Start

One-Command Setup

git clone https://github.com/morningstarnasser/MORNINGSTAR-Vision-AI.git
cd MORNINGSTAR-Vision-AI
chmod +x setup_and_benchmark.sh
./setup_and_benchmark.sh

This pulls all base models, creates all 3 variants, and runs the full benchmark suite automatically.

Manual Setup

# Install Ollama (if not installed)
curl -fsSL https://ollama.com/install.sh | sh

# Clone & build
git clone https://github.com/morningstarnasser/MORNINGSTAR-Vision-AI.git
cd MORNINGSTAR-Vision-AI

# Create models (choose what you need)
ollama create morningstar -f Modelfile              # 14B — Fast & powerful
ollama create morningstar-32b -f Modelfile.32b      # 32B — Maximum quality
ollama create morningstar-vision -f Modelfile.vision # Vision — See images

# Run
ollama run morningstar

System Requirements

Model	RAM	Disk	GPU
morningstar (14B)	16 GB+	10 GB	Optional
morningstar-32b	24 GB+	20 GB	Optional
morningstar-vision	16 GB+	8 GB	Optional

Apple Silicon (M1/M2/M3/M4), NVIDIA CUDA, and AMD ROCm all supported. CPU-only works too.

---

Usage

Chat Mode

ollama run morningstar
>>> Write a REST API in Rust with Axum and JWT auth

API Mode

curl http://localhost:11434/api/generate -d '{
  "model": "morningstar",
  "prompt": "Write a React component with TypeScript generics",
  "stream": false
}'

Python

import requests

response = requests.post("http://localhost:11434/api/generate", json={
    "model": "morningstar-32b",
    "prompt": "Design a microservices architecture for an e-commerce platform",
    "stream": False,
})
print(response.json()["response"])

Vision (Image Analysis)

ollama run morningstar-vision
>>> [Upload an image] What's in this image? Write code to recreate it.

Morningstar CLI

npm i -g morningstar-cli
morningstar --model morningstar

---

The Three Models

Download from HuggingFace

Model	Parameters	Size	HuggingFace
morningstar	14.2B	9 GB	kurdman991/morningstar-14b
morningstar-32b	32.5B	20 GB	kurdman991/morningstar-32b
morningstar-vision	13B	8 GB	kurdman991/morningstar-vision

morningstar — The Default (14B)

Base	Qwen2.5-Coder-14B-Instruct
Parameters	14.2 Billion
Context	128K tokens
Size	~9 GB
Best For	Daily coding, fast iteration
HuggingFace	kurdman991/morningstar-14b

morningstar-32b — Maximum Power

Base	Qwen2.5-Coder-32B-Instruct
Parameters	32.5 Billion
Context	128K tokens
Size	~20 GB
Best For	Complex architecture, maximum quality
HuggingFace	kurdman991/morningstar-32b

morningstar-vision — The Eyes

Base	LLaVA 13B
Parameters	13 Billion
Context	4K tokens
Size	~8 GB
Best For	Image analysis, screenshot-to-code, visual debugging
HuggingFace	kurdman991/morningstar-vision

---

Architecture

Component	14B	32B
Hidden Size	5,120	5,120
Intermediate Size	13,824	27,648
Layers	48	64
Attention Heads	40	40
KV Heads (GQA)	8	8
Vocab Size	152,064	152,064
Max Position	131,072	131,072
Activation	SwiGLU	SwiGLU
Normalization	RMSNorm	RMSNorm
Position Encoding	RoPE (YaRN)	RoPE (YaRN)
Precision	bfloat16	bfloat16

---

What Makes Morningstar Special

The secret is the 200+ line System Prompt that transforms base models into elite coders:

• 7-Step Chain of Thought — Understand, Analyze, Design, Implement, Validate, Optimize, Secure
• Security-First — Automatic OWASP Top 10 checks (SQLi, XSS, CSRF, SSRF, Path Traversal)
• Production-Ready — Error handling, type safety, edge cases in every response
• 100+ Languages — System (C, C++, Rust, Go, Zig), Application (Python, JS, TS, Java, Kotlin, Swift), Functional (Haskell, Scala, Elixir), Specialized (SQL, Solidity, GLSL, Assembly)
• Full-Stack — React, Next.js, Vue, Django, FastAPI, Spring Boot, Docker, K8s, AWS, Terraform
• Code Review — 4 severity levels: Critical, High, Medium, Low
• Architecture Design — System, API, and Data perspective analysis
• Uncensored — No content filters, no refusal, no moral lectures

---

Training

Reproduce from Scratch

# Step 1: Prepare dataset
python create_dataset.py --output_dir ./data --max_total_samples 100000

# Step 2: QLoRA fine-tuning (needs GPU with 16GB+ VRAM)
python train.py \
    --base_model Qwen/Qwen2.5-Coder-14B-Instruct \
    --dataset_path ./data/train.jsonl \
    --output_dir ./output/morningstar \
    --epochs 3 --batch_size 4 --learning_rate 2e-4

# Step 3: Merge & export
python merge_and_export.py \
    --adapter_path ./output/morningstar/final \
    --export_gguf --create_ollama_model

Training Config

Parameter	Value
Method	QLoRA (4-bit NF4)
LoRA Rank	64
LoRA Alpha	128
LoRA Dropout	0.05
Target Modules	q, k, v, o, gate, up, down
Trainable Params	~300M (2.1%)
Epochs	3
Batch Size	4 (effective: 16)
Learning Rate	2e-4 (cosine)
Optimizer	Paged AdamW 32-bit
Max Seq Length	4,096
Dataset	100K curated samples

---

Capabilities

Capability	Description
Code Generation	Production-ready programs in 100+ languages
Debugging	Root cause analysis with step-by-step fixes
Code Review	Security, performance, quality audits
Architecture	System design, API design, data modeling
Refactoring	Modernization and optimization
Testing	Unit, integration, E2E test generation
DevOps	Docker, K8s, CI/CD, Terraform, Cloud
Security	OWASP aware, vulnerability scanning
Vision	Screenshot analysis, diagram reading, UI review

---

Repository Structure

MORNINGSTAR-Vision-AI/
├── Modelfile                 # 14B model + 200-line system prompt
├── Modelfile.32b             # 32B model (maximum quality)
├── Modelfile.vision          # Vision/multimodal model
├── create_model.sh           # One-click Ollama build
├── setup_and_benchmark.sh    # Full auto-setup + benchmark
├── benchmark.sh              # 19-test benchmark suite
├── train.py                  # QLoRA fine-tuning
├── create_dataset.py         # Dataset preparation
├── merge_and_export.py       # LoRA merge + GGUF export
├── example_usage.py          # API usage demos
├── config.json               # Model architecture config
├── tokenizer_config.json     # Tokenizer + ChatML template
├── requirements.txt          # Python dependencies
└── README.md

---

Run the Benchmark

chmod +x benchmark.sh
./benchmark.sh morningstar
./benchmark.sh morningstar-32b

Tests include: Python, JavaScript, TypeScript, Rust, Go, C, SQL, React, Docker, Regex, algorithms, data structures, and more.

---

Examples

Rust — Async REST API with Axum

> Write an async REST API in Rust with Axum, JWT auth, and PostgreSQL

Morningstar generates a complete project with proper error handling, middleware, connection pooling, and security headers.

Python — Distributed Task Queue

> Build a distributed task queue with Redis and asyncio

Full implementation with worker processes, retry logic, dead letter queues, and monitoring.

TypeScript — React Component Library

> Create a type-safe Button component with variants, sizes, and compound patterns

Generates fully typed components with generics, proper forwarding, and Storybook stories.

Vision — Screenshot to Code

> [Upload screenshot] Recreate this UI in React with Tailwind CSS

Morningstar-Vision analyzes the image and generates pixel-accurate React code.

---

License

Apache License 2.0 — Free for commercial and personal use. No restrictions.

---

Citation

@misc{morningstar-vision-ai-2025,
  title   = {Morningstar Vision AI: Elite Open-Source Coding & Vision System},
  author  = {Ali Nasser},
  year    = {2025},
  url     = {https://github.com/morningstarnasser/MORNINGSTAR-Vision-AI}
}

---

The God of Code. Now with Eyes. Open Source. Forever Free.

_{If this project helped you, consider giving it a star.}