supra-nexus-o1-instruct-mlx-4bit

4-bit quantized MLX format of the instruction-tuned Supra Nexus O1 model for efficient inference on Apple Silicon.

🔗 Model Collection

Base Models

• 🤖 supra-nexus-o1-instruct - Instruction-following model
• 💭 supra-nexus-o1-thinking - Chain-of-thought reasoning model

Available Formats

Instruction Model

• 📦 GGUF | 🍎 MLX | ⚡ MLX 4-bit

Thinking Model

• 📦 GGUF | 🍎 MLX | ⚡ MLX 4-bit

Training Data

• 📊 supra-nexus-o1-training - Complete training dataset

💡 Key Features

• Transparent Reasoning: Shows thought process using <thinking> tags
• Chain-of-Thought: Step-by-step problem solving approach
• Self-Improvement: Trained with recursive improvement examples
• Multi-Format: Available in multiple formats for different platforms

🚀 Quick Start

Using with MLX (4-bit Quantized)

from mlx_lm import load, generate

# Load 4-bit quantized model (75% smaller)
model, tokenizer = load("Supra-Nexus/supra-nexus-o1-instruct-mlx-4bit")

# Generate with chain-of-thought
prompt = "Solve step by step: What is 25% of 480?"
response = generate(model, tokenizer, prompt=prompt, max_tokens=500)
print(response)

Benefits of 4-bit Quantization

• 🚀 75% smaller model size
• ⚡ Faster inference on M1/M2/M3 Macs
• 💾 Lower memory requirements
• ✨ Minimal quality loss

📈 Performance

The O1 models excel at:

• Complex reasoning tasks
• Step-by-step problem solving
• Mathematical computations
• Code generation and debugging
• Creative writing with logical structure

🏗️ Architecture

Based on Qwen2.5 architecture with:

• Custom fine-tuning for reasoning
• Chain-of-thought training
• Self-improvement capabilities
• Identity preservation techniques

🔬 Training Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Training Framework: Zoo Gym
• Dataset: supra-nexus-o1-training
• Training Duration: Multiple iterations with self-improvement
• Hardware: NVIDIA A100 GPUs

📚 Resources

• 📖 GitHub Repository - Source code and documentation
• 🏢 Supra Foundation - Organization behind O1
• 🐦 Twitter - Latest updates
• 💬 Discord - Community support

📄 Citation

@software{supra_nexus_o1_2025,
  title = {Supra Nexus O1: Advanced Reasoning Models},
  author = {Supra Foundation},
  year = {2025},
  url = {https://github.com/Supra-Nexus/o1}
}

📝 License

Apache 2.0 - See LICENSE for details.

---

Building transparent AI reasoning systems 🧠✨