An open-weight model is an AI system whose trained parameter values — the numerical weights learned during training — are publicly released so that anyone can download, inspect, run, and modify the model without going through an API. Open-weight is often distinguished from fully open-source: the weights may be available while the training code, dataset, or data curation pipeline remains proprietary. Meta's Llama series and Mistral AI's models are prominent examples where weights are downloadable but training data details are not fully disclosed.

Think of it like receiving a finished, printed circuit board with the schematic available but the manufacturing process kept private. You can probe every trace, solder on new components, and repurpose the board for a different device — you just cannot reproduce the factory line that produced it.

Open-weight access enables workflows that are impossible with closed API-only models. For instance, a medical AI team can load Llama 3 70B, fine-tune it on de-identified patient notes that cannot leave their servers, and serve predictions entirely on-premise:

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-70B")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-70B")

The weights can also be quantized to reduce memory footprint, merged with LoRA adapters after domain-specific fine-tuning, or distilled into a smaller model for edge deployment.

Open-weight models are a subset of the broader Foundation Model category, and they are the starting point for most Fine-tuning and Knowledge Distillation workflows. Quantization techniques are especially important here because most open-weight models are too large to run in full precision on consumer hardware.

You will encounter the open-weight distinction most often when evaluating self-hosting options, comparing privacy and compliance trade-offs between local and API-based deployment, or reading licensing terms that govern commercial use of a given model.