Open-Source AI refers to models where the underlying code, architecture, and training data are publicly available, though the industry term has largely shifted
Open-source AI spans open-weight models (downloadable parameters: Llama, Mistral, Qwen families), open training/inference frameworks (PyTorch, vLLM), and open datasets. Teams self-host or use managed inference, customize via fine-tuning, and accept responsibility for safety, updates, and ops that API providers otherwise carry.
Open models are the counterweight to API concentration: they enable data sovereignty, on-prem and edge deployment, deep customization, and dramatic unit-cost reduction at scale. The 2026 frontier gap has narrowed enough that tuned open models win many production workloads: making build-vs-API a genuine architectural decision.
Often not strictly: many ship weights under licenses with usage restrictions, without training data or full recipes. 'Open-weight' is the precise term; read licenses before commercial deployment.
On well-scoped tasks at volume (cost), data-sensitive workloads (control), latency-critical paths, and customization-heavy products. Frontier APIs retain the edge on hardest reasoning and zero-ops simplicity: hybrids are the norm.
GPU capacity (owned or cloud), an inference server (e.g. vLLM), monitoring, security patching, and eval discipline for model updates: real ops work that should be priced into any build-vs-buy comparison.