What is an open-weight model?
An open-weight model is an AI model whose trained weights, the billions of numbers that make up the model, are released for anyone to download and run. The publisher usually provides the model architecture and inference code as well, so the model can be loaded on a buyer’s own GPUs with standard open-source inference servers. What the publisher typically keeps back is the training data and the full training recipe, and the license may impose conditions. Meta’s Llama family is the best-known example.
This makes open-weight different from fully open source. An open-source model publishes weights, code, and training data under a license with no field-of-use restrictions. An open-weight model gives a company control over where the model runs and what happens to its data, but not full visibility into how the model was built, and the license may limit commercial use above a certain scale, forbid specific applications, or require attribution.
Why open-weight models matter
Open-weight models are what make private AI practical. A company can download a model that competes with hosted services, run it inside its own network or a dedicated cloud tenancy, and never send a prompt to a third party. The cost moves from a per-token bill to GPU capacity the company owns or rents by the hour. The weights can be fine-tuned, quantized to fit smaller hardware, and pinned to a version so that behavior does not change until the company decides to change it. FlatClaw, the open-source private AI platform from Kirk Tech Solutions, is built around running open-weight models on infrastructure the client controls.
Open-weight models in practice
A regional bank wants an assistant that summarizes loan files and drafts credit memos. Its policy forbids customer financial data leaving its environment. The bank deploys a 30-billion-parameter open-weight model on a single GPU server in its data center, quantized to 8-bit so the weights fit in memory with room for concurrent users. Risk and compliance review the model card and license before approval, and the model version is recorded in the bank’s model inventory.
What to watch for
- Read the license for user thresholds, prohibited uses, and attribution requirements before building on the model.
- Treat the model card as the main source of information about training data, and note what it leaves out.
- Pin the exact weight files and verify their checksums, because publishers do update releases.
- Size GPU memory for the weights, the KV cache, and the number of simultaneous users, since all three compete for the same space.