Model Weights

Model weights are the concrete, saved numeric values of a trained neural network's parameters — the actual file (or set of files) you'd download, load into memory, and run to use a model, as opposed to the abstract architecture or the training process that produced them. When people distinguish between "open-weight" and "closed" models, this is precisely the distinction: open-weight models (like Llama, Mistral, Qwen, and DeepSeek) publish their trained weight files for anyone to download, inspect, self-host, and fine-tune; closed models (like Claude, GPT, and Gemini) keep their weights private on the provider's own infrastructure, accessible only through an API, with no way to download or directly inspect the underlying numbers. This distinction matters enormously for SaaS and AI builders as a genuine architectural and business decision, not just a philosophical one. Open weights enable: full data sovereignty (running the model entirely within your own infrastructure, critical for regulated industries or strict privacy requirements), unlimited fine-tuning and customization without a provider's API constraints, no per-token API cost once infrastructure is provisioned (though hardware/hosting costs replace it), and no dependency on a third party's uptime, pricing changes, or policy shifts. Closed models typically offer: no infrastructure/GPU management burden, access to the most capable frontier-level models (open-weight models have historically lagged 6-18 months behind the closed frontier, though the gap narrows over time), simpler pay-per-use economics at low-to-moderate volume, and automatic access to model improvements without any migration work. A concrete worked example: a healthcare SaaS company handling sensitive patient data under strict compliance requirements decides that sending any data to a third-party API is a non-starter; they download Llama's open weights, self-host the model on their own GPU infrastructure inside their compliance boundary, and accept a somewhat less capable model in exchange for full data control — a trade-off a general marketing SaaS tool with no such constraint would likely make the opposite way, favoring a closed frontier API for simplicity and capability. A middle-ground option worth knowing exists between fully closed and fully open: some providers offer "open-weight, restricted-license" models (weights are downloadable and self-hostable, but usage is bound by a license restricting certain commercial uses or requiring attribution above a usage threshold) — meaning "open-weight" alone doesn't automatically mean unrestricted commercial use, and builders should read a model's specific license terms before assuming self-hosting removes all legal/commercial constraints, since license terms vary significantly between model families and can change between versions.