Open Weights

Open weights refers to AI models where the trained parameters — the numerical values that define the model's knowledge and behaviour — are publicly available for download. This allows anyone to run the model on their own hardware without relying on a cloud API.

Open weights vs open source

The distinction matters. True open source means releasing not only the model weights but also the complete training code, training data, and documentation — everything needed to reproduce the model from scratch. Most "open" AI models are open-weight but not fully open source. Meta's Llama, Mistral's models, and Google's Gemma release weights but not complete training pipelines or data.

Why open weights matter

• Independence: You can run the model on your own infrastructure without depending on an API provider. No risk of service changes, price increases, or discontinuation.
• Privacy: Data never leaves your environment. Sensitive documents, customer data, and proprietary information are never sent to a third party.
• Customization: Open-weight models can be fine-tuned on your specific data for better performance on your tasks.
• Cost at scale: For high-volume applications, running your own model can be significantly cheaper than paying per-token API prices.
• Research and innovation: Researchers can study, modify, and improve the model, accelerating the field.

Popular open-weight models

• Llama (Meta): One of the most capable open-weight model families, available in multiple sizes.
• Mistral and Mixtral (Mistral AI): Efficient models known for strong performance relative to their size.
• Gemma (Google): Lightweight models optimized for accessibility and efficiency.
• Qwen (Alibaba): Multilingual models with strong performance across languages.

Licensing considerations

Open-weight models come with various licenses that govern acceptable use. Some are permissive (use for any purpose including commercial), while others restrict certain uses (military applications, generating disinformation). Always review the license before deploying an open-weight model in production.

Trade-offs

Open-weight models require technical expertise to deploy and maintain. You manage infrastructure, updates, and security. API-based models handle all of this for you but at the cost of dependency and potentially higher per-query costs at scale.