What are open-weight AI models and why do companies care?

Open-weight AI models publish their trained parameters — the billions of numerical weights that encode what the model learned — so anyone can download, run, fine-tune, and inspect them. OpenAI’s gpt-oss release is the clearest recent example: the weights are downloadable and runnable on your own hardware, but not served through OpenAI’s API or ChatGPT. The category also includes Meta’s Llama family, Mistral, Google’s Gemma, Alibaba’s Qwen, and DeepSeek. Licenses vary from permissive (Apache 2.0, MIT) to custom terms with use restrictions, so read the specific license before shipping.

Why companies care comes down to control versus convenience. Closed models are rented by the token; open-weight models are hosted by you. In exchange for that work you get data residency (inputs never leave your infrastructure — often non-negotiable in healthcare, finance, and legal), fine-tuning on proprietary data, lower latency, more predictable cost at high volume, and vendor independence if a provider changes pricing or deprecates a model.

The trade-offs are real: you inherit the GPU bill, the serving and monitoring burden, and the safety guardrails that closed APIs bundle for you. Managed inference providers like Together, Fireworks, and AWS Bedrock sit between the two — open-weight models with API-style access.

For most small businesses the API path is faster; self-hosting earns its keep when volume, regulation, or proprietary data make control worth the overhead.