Open-Weight Model

An open-weight model is a machine learning model whose trained weights (its learned numeric parameters) are released publicly so you can download and run it on your own hardware. It is not automatically open source: the training data, training code and licence terms may be withheld or restricted.

Why it matters

Having the weights means you can host the model yourself, keep prompts and data on your own servers, fine-tune it for a narrow task, and avoid per-request pricing from a hosted API. The trade-off is that you carry the operational burden: GPUs, serving software, updates and safety filtering. Licences differ widely, and some limit commercial use, large deployments or specific uses, so the licence text matters as much as the download link.

Example

# Run a downloaded model locally with a typical inference tool
model = load("my-model-8b-instruct")
print(model.generate("Summarise this log file: ..."))

An 8-billion-parameter model stored in 16-bit precision needs roughly 16 GB of memory for the weights alone, which is why people often apply quantization.

Common misconceptions

• "Open weights means open source." Open source implies a licence that permits free use, modification and redistribution, plus access to the source. Many weight releases do not meet that bar.
• "You can reproduce the model." Without the data and training recipe you can only use or fine-tune it.
• "Self-hosting is always cheaper." Idle GPUs and engineering time can outweigh API fees.

Related

• Open source
• Quantization
• Fine-tuning
• LLM

This page is general information written by our editorial team and updated when documentation changes. Servers and providers differ, so test changes safely first. See our editorial policy and disclaimer.