Nanochat - The $100 ChatGPT clone

Nanochat is a versatile AI assistant that can help with a wide variety of tasks. You can use it for general conversation, asking questions about various topics, getting help with homework or research, brainstorming ideas, learning new concepts, practicing language skills, getting coding assistance, and much more. While it's designed as an educational demonstration, it provides genuinely useful responses across many domains including science, mathematics, history, literature, technology, and everyday problem-solving.

Yes, Nanochat is completely free to use with no hidden costs, subscriptions, or usage limits. You don't need to create an account, provide payment information, or download any software. Simply start chatting directly on this page. The project is open source and funded by the AI research community's commitment to democratizing access to artificial intelligence technology.

Nanochat is designed as a lightweight, educational AI model rather than a commercial product. While larger models like GPT-4 or Claude have more parameters and broader capabilities, Nanochat demonstrates impressive performance for its size. With 560 million parameters trained on high-quality educational data, it achieves competitive scores on standard benchmarks. The key advantage of Nanochat is its accessibility - the complete training process costs just $100 and takes only 3-4 hours, making it perfect for learning how modern language models work. It's particularly strong in educational contexts, clear explanations, and demonstrating fundamental AI capabilities.

Absolutely! That's one of the primary goals of the Nanochat project. The complete source code, training scripts, and documentation are available under an MIT open source license. You can download everything you need, follow the step-by-step training guide, and create your own customized version. The process requires access to GPUs (8×H100 or equivalent) and costs approximately $100 for the full training pipeline. Many cloud providers offer GPU rentals at hourly rates, making this accessible to individuals and small organizations. The project includes detailed instructions for environment setup, data preparation, tokenizer training, pre-training, fine-tuning, and evaluation.

Nanochat's training methodology is remarkable for its efficiency and transparency. The project demonstrates that effective language models don't require massive corporate resources. Key innovations include:

• Efficient scaling: Optimal data-to-parameter ratio following Chinchilla scaling laws
• Quality over quantity: Training on curated FineWeb-EDU dataset rather than raw web scrapes
• Custom tooling: Purpose-built Rust tokenizer for training efficiency
• Complete pipeline: From tokenizer training through reinforcement learning
• Reproducibility: Every step documented with exact configurations and hyperparameters

This approach proves that thoughtful engineering and quality data can achieve impressive results even with limited resources.

As an educational demonstration model, Nanochat has some limitations compared to larger commercial systems. It may occasionally produce inaccurate information, has a knowledge cutoff based on its training data, cannot access real-time information or browse the internet, and may struggle with very specialized or technical queries outside its training domain. The model has 560 million parameters, which is significantly smaller than frontier models, so its reasoning capabilities are more limited. However, these limitations are actually valuable for educational purposes, as they help users understand the current state and constraints of AI technology. Nanochat excels at demonstrating core language model capabilities and serving as a learning tool for AI education.

Nanochat was created as an open-source educational project to demonstrate that modern AI technology is accessible to everyone, not just large tech companies. The project aims to demystify large language model training by providing a complete, reproducible example that costs just $100. By making every aspect of the training process transparent and well-documented, the project serves as an invaluable learning resource for students, researchers, and developers worldwide. The goal is to advance AI education and enable more people to understand, experiment with, and contribute to the field of artificial intelligence.

Your privacy is important. When you use Nanochat, your conversations are processed to generate responses but are not permanently stored or used for additional training. The chat interface operates through a secure connection, and no personal information is required or collected to use the service. As with any AI system, you should avoid sharing sensitive personal information, passwords, or confidential data in your conversations. The open-source nature of the project means you can review the code to understand exactly how data is handled, and you can even deploy your own private instance if you need guaranteed data isolation.

Yes! Nanochat is released under the MIT License, one of the most permissive open-source licenses. This means you can use the code, model, and training methodology for commercial purposes, modify it to suit your needs, integrate it into your products, and even create derivative works. The only requirement is that you include the original license text in your distribution. This makes Nanochat an excellent starting point for businesses wanting to build custom AI solutions, researchers developing new techniques, or developers creating AI-powered applications without licensing restrictions or ongoing fees.

Training your own Nanochat model requires several components:

• Hardware: Access to 8 H100 GPUs (or equivalent) for approximately 3-4 hours. Cloud GPU rental services make this accessible.
• Budget: Approximately $100 for GPU time, dataset storage, and related costs.
• Software: Python environment with PyTorch, Rust compiler for the tokenizer, and standard ML libraries.
• Data: The FineWeb-EDU dataset (automatically downloaded by the training scripts).
• Knowledge: Basic understanding of Python and command-line operations. The detailed documentation guides you through each step.

The complete training pipeline includes tokenizer training (~1 minute), pre-training (~3 hours), mid-training (~7 minutes), and supervised fine-tuning (~7 minutes). All scripts are provided and well-documented.