The vocabulary of AI has blurred to the point of being misleading. Three labels are used interchangeably: closed, open-weight, and open-source, yet they describe genuinely different bundles of artifacts and rights, with direct impact on cost, control, and risk. Openness is not a binary. A model is assembled from separable parts like the trained weights, the training data, the training code and recipe, and the license governing all of it, and a provider can release any subset of them. The weights are the heart of it: the billions of numerical parameters a model adjusts during training, the learned values that determine how it turns an input into an output. Knowing which of these parts you are actually getting is the whole game.
Closed Model
A closed model releases none of them. The weights never leave the provider's infrastructure, and capability is metered through an API; you rent intelligence rather than possess it. OpenAI's GPT-5 series, Anthropic's Claude, and Google's Gemini are the canonical examples, and they remain the frontier. You get the most capable models with no infrastructure to run and updates managed for you, but your data leaves your environment, pricing and availability sit at the provider's discretion, you cannot inspect or fine-tune the model, and a version you depend on can change beneath you. You are buying capability and ceding control.
Open-Weight Model
An open-weight model publishes those trained parameters for anyone to download, run, and fine-tune, while keeping the training data, code, and full recipe private. This is the fastest-growing category. Meta's Llama, Mistral, Alibaba's Qwen, DeepSeek, and Google's Gemma live here (as does OpenAI's own gpt-oss, released in August 2025 under a permissive Apache 2.0 license, its first open-weight model since GPT-2). The appeal is control without building from scratch: run the model on your own hardware, keep data in your environment, and fine-tune it on proprietary information to embed organizational intelligence into a model you actually hold (we wrote about organizational intelligence here Organizational Intelligence Embedded in Fine-Tuned LLMs). What you give up is reproducibility and often true freedom because you cannot audit how the model was built, data provenance is opaque, and many of these licenses carry usage restrictions that would disqualify them as open-source software.
Open-Source Model
An open-source model, strictly defined, releases everything needed to rebuild it. The Open Source Initiative's Open Source AI Definition, published at the end of 2024, requires not just the freedom to use, study, modify, and share a model but the training code, a description of the data, and enough detail to reconstruct an equivalent system. By that bar, most models marketed as open are merely open-weight. The genuine article is rare: the Allen Institute for AI's OLMo family, now in its third generation, ships the weights alongside its Dolma training corpus, the code, the checkpoints, and the logs. The payoff is full transparency and reproducibility, the ability to audit for bias and safety, and no dependence on any vendor. The cost is capability, as fully open models still trail the frontier.
The three labels, then, are points on a spectrum, each trading one thing for another: closed trades control for capability, open-weight trades reproducibility for ownership of the runtime and the customization layer, open-source trades some capability for total transparency. The strategic question for an enterprise is not which is best but which layer of the stack it needs to own. The direction of travel is clear: open-weight models now account for roughly a third of the tokens served across major model-routing platforms, and the arrival of frontier labs in the category is accelerating the shift (continuing to push for AI as an Operating System).
Alten Capital invests in exceptional management teams to accelerate high-growth technology services businesses. Reach out to explore partnership opportunities.