Open-source models - when they actually matter
Open-source (or "open-weight") models are ones whose parameters are publicly downloadable, letting anyone run them on their own hardware instead of calling a company's API. Meta's Llama, Mistral, DeepSeek, and Alibaba's Qwen are the names you'll hear most often. They matter for privacy-sensitive use cases, cost control at huge scale, or full customization - not usually for quality, since top closed models are typically still ahead.
Running an open model yourself means you own infrastructure costs (GPUs, hosting) and lose the "it just works" reliability of a managed API - a real tradeoff, not a free lunch. Companies choose open-source mainly for data sovereignty (nothing leaves their servers), predictable fixed cost at extreme scale, or the ability to fine-tune deeply for a narrow task.
Advisory case: if a healthcare-AI conversation ever raises "patient data can't leave our servers," that's exactly the scenario where open-source/self-hosted becomes the right call over a closed API - a good real test of whether you can apply this distinction correctly.
When do open-source (open-weight) models matter most?