README.md

metadata

title: Opus Research
emoji: 🧠
colorFrom: purple
colorTo: blue
sdk: static
pinned: false

🧠 Opus Research

Training AI models from scratch, one parameter at a time.

"We stand at the right place at the right time."

👋 About Us

We're two teenage AI enthusiasts (ages 15 & 17) passionate about understanding AI from the ground up. Instead of just using pre-trained models, we build them ourselves.

🚀 Our Models

Model	Parameters	Architecture	Status
Opus 1.5	0.88B	LLaMA-style	✅ Released
Opus 2.0	3B+	LLaMA + Reasoning	🔜 Coming Soon

🔬 Research Focus

• Training from scratch - No pre-trained weights, 100% original
• Chain-of-thought reasoning - Teaching models to think before answering
• Efficient architectures - Sub-3B models that run on consumer GPUs

📊 Opus 1.5 Highlights

• 0.88 billion parameters
• ~2 GB VRAM for inference
• 42 hours training on 2x RTX 4090
• LLaMA architecture with RoPE, SwiGLU, GQA, FlashAttention-2

🔗 Links

• 📦 Opus 1.5 on HuggingFace

⚡ Powered By

• ❤️ Two teenagers who should probably be doing homework
• 🎮 Two RTX 4090s that haven't known rest since December 2025
• 🔥 Questionable GPU thermals
• 📊 A concerning amount of time staring at loss curves
• 💡 48GB of VRAM
• ⚡ ~847 kWh of electricity we'll never get back
• 💸 Our parents' electricity bill
• 🤔 The firm belief that we could've just used ChatGPT

But where's the fun in that?