Open Weights
AI models whose trained parameters (weights) are publicly released, allowing anyone to download, run, modify, and fine-tune the model locally.
What Is Open Weights?
Open-weights AI models release the trained numerical parameters of the neural network publicly, typically on platforms like Hugging Face. This is distinct from fully "open source" in the traditional sense — the weights are shared, but the training data, training code, or commercial use may be restricted depending on the licence. Examples include Meta's Llama series, Mistral, Qwen, and Falcon.
Open-weights models can be run entirely on local hardware (laptop, workstation, data centre) with no API calls or usage fees. Tools like Ollama, LM Studio, and llama.cpp make running 7B–70B parameter models on consumer hardware practical. For production, quantised versions (4-bit or 8-bit) reduce memory requirements significantly.
Why It Matters
Open-weights models enable a category of AI use cases impossible with closed models: fully private inference (no data leaves your hardware), unlimited customisation, fine-tuning on proprietary data without sharing it with a third party, and cost-free inference at scale. The gap between open-weights and frontier closed models has narrowed dramatically — Llama 3.1 405B and Qwen 2.5 72B rival GPT-4 on many benchmarks.
Real-World Examples
Local Privacy-First AI
A healthcare company runs Llama 3 locally so patient data never leaves their secure environment, complying with HIPAA without expensive enterprise API agreements.
Fine-Tuned Specialist
A developer fine-tunes Mistral 7B on 1,000 examples of their company's SQL patterns, creating a cheap, fast, specialised SQL assistant.