Open-Source AI

Open-source AI refers to AI models, frameworks, and tools that are made freely available for anyone to use, inspect, modify, and distribute. Prominent examples include Meta's Llama models, Stability AI's Stable Diffusion, and the Hugging Face ecosystem.

What "open source" means in AI

The term is used with varying degrees of openness:

• Fully open: Model weights, training code, training data, and evaluation code are all publicly available. Rare in practice.
• Open weights: The trained model weights are available for download and use. Training data and code may not be shared. This is the most common form (Llama, Mistral, Gemma).
• Open API: The model is accessible via a free API but weights are not downloadable. Not truly open source but sometimes marketed as such.

Advantages of open-source AI

• Control: You can run models on your own infrastructure, keeping data private and avoiding vendor lock-in.
• Customisation: You can fine-tune, modify, and combine models for your specific needs.
• Cost: No per-query API fees — you pay only for the compute to run the model.
• Transparency: You can inspect the model to understand its behaviour and biases.
• Community: Large communities contribute improvements, tools, and knowledge.

Leading open-source models

• Llama (Meta): Strong general-purpose models in various sizes.
• Mistral / Mixtral: Efficient European models with strong performance.
• Gemma (Google): Lightweight models for on-device and cloud use.
• Qwen (Alibaba): Competitive multilingual models.
• Stable Diffusion: The dominant open-source image generation model.

Open source vs proprietary: the trade-offs

Proprietary models (ChatGPT, Claude) typically offer higher peak performance, managed infrastructure, and regular updates. Open-source models offer control, customisation, and potential cost savings at scale but require more technical expertise to deploy and maintain.

Licensing nuances

Not all "open" models use standard open-source licences. Many use custom licences with restrictions — for example, prohibiting commercial use above a certain user threshold or requiring attribution. Always read the licence before deploying.