Open Weights

Model weights released so others can download, run, and often fine-tune them locally or on their own cloud, usually under a permissive or research license.

How it works

When a lab releases open weights, they publish the trained model parameters — typically as a set of binary files in formats like safetensors or GGUF — along with tokeniser files and model configuration. Users can download these weights and load them into inference frameworks (llama.cpp, vLLM, Transformers) to run locally. Open weights differ from fully open-source AI: the weights are the output of training, but the training data, code, and detailed training procedures may not be published.

Why it matters

Open weights have been the single most powerful force for democratising access to advanced AI. LLaMA, Mistral, DeepSeek, and Kimi K2 open weights enabled thousands of fine-tuned variants, independent evaluation, safety auditing, and deployment in resource-constrained environments without API costs. They also enable local, private inference — critical for regulated industries. The debate over open vs. closed weights is one of the defining policy questions in AI governance, with open weights proponents arguing for innovation and access, opponents citing misuse risks.