Open-Weight Models

Open-weight models are AI models whose trained parameters (weights) are publicly released for download and use. This is distinct from truly open-source models, where the training data, training code, and evaluation methodology are also made available. The distinction matters because it affects what you can do with the model and how much you can trust it.

Open-weight versus open-source

The AI industry uses these terms loosely, but the distinction is important:

• Open-weight: The trained model is available for download and use. You can run it, fine-tune it, and (depending on the licence) deploy it commercially. However, you cannot see exactly what data it was trained on or reproduce the training process.
• Open-source (in the traditional software sense): The training code, training data, data processing pipeline, and evaluation code are all available. Anyone can reproduce the model from scratch.
• Proprietary: The model is available only through an API. You cannot download, inspect, or self-host it.

Most models marketed as "open-source" are actually open-weight. Meta's Llama, Mistral's models, and Google's Gemma release weights but not complete training data or code.

Why the distinction matters

• Reproducibility: Without training data and code, you cannot verify the model's behaviour or reproduce it independently.
• Bias auditing: Without knowing the training data, it is harder to assess potential biases and limitations.
• Regulatory compliance: Some regulations may require transparency about training data, which open-weight alone does not provide.
• Licensing: Open-weight models come with various licences, some more restrictive than others. Always read the licence carefully.

Prominent open-weight models

• Llama 3 (Meta): Available in various sizes (8B to 405B parameters). Commercial use permitted with some restrictions.
• Mistral / Mixtral (Mistral AI): Efficient European models with strong multilingual capabilities.
• Gemma (Google): Smaller, efficient models from Google's DeepMind team.
• Qwen (Alibaba): Strong models from Alibaba Cloud, particularly good for multilingual tasks.
• Phi (Microsoft): Small, efficient models optimised for specific reasoning tasks.
• Command R (Cohere): Optimised for RAG and enterprise tasks.

Benefits for enterprises

• Data privacy: Run models on your own infrastructure with no data leaving your environment.
• Cost predictability: After the initial hardware investment, inference costs are fixed.
• Customisation: Fine-tune on your specific data and use cases.
• No vendor lock-in: Switch between models without changing your infrastructure.
• Compliance: Demonstrate to regulators exactly what model you are running and how data is handled.

Challenges

• Infrastructure: Self-hosting requires GPU infrastructure and operational expertise.
• Updates: You are responsible for updating to newer model versions.
• Support: No vendor support — you rely on community resources and your own team.
• Performance gap: Open-weight models generally trail the best proprietary models (Claude, GPT-4) on the hardest tasks, though the gap is narrowing.

The strategic calculus

For most enterprises, the optimal strategy is a mix: use proprietary APIs for tasks requiring the best available quality, and use open-weight models for high-volume, cost-sensitive, or privacy-critical tasks.