# Open Weights An open-weight model is one whose trained parameters (the weights file) are published for anyone to download, run and modify. The weights are the model: billions of numbers that encode everything it learned. With the file in hand you can run inference on your own hardware, fine-tune it, or serve it to customers. ## Open weights are not open source "Open source" in the strict sense means you can reproduce the thing: code, data and training recipe. Most open-weight releases publish only the finished weights plus inference code, under a licence that ranges from permissive (Apache 2.0, MIT) to custom terms with usage conditions. You get the artefact, not the factory. See [[Open source is often misunderstood]]. ## Properties that matter strategically - **Irrevocable.** Once released, weights cannot be recalled. Any policy applied afterwards only affects future releases. - **No counterparty.** There is no vendor contract, no API account, no one to audit. This is why they route around enterprise procurement: vendor-risk reviews and country-of-origin screening assume someone is on the other side of the table. - **Portable.** The same file runs on a hyperscaler, a [[Neoclouds|neocloud]] or a laptop, so value moves to whoever serves the tokens ([[Token Economics of Inference]]). - **Composable.** Companies [[Pretraining, Post-Training and Fine-Tuning|post-train]] on top of them and ship the result as their own product. > [!important] The procurement point > Procurement controls vendors. Open weights have no vendor. The control surface that remains is where and how the model is run: see [[Jurisdiction and Control Planes]]. Examples referenced in the article: Llama (Meta), Qwen (Alibaba), DeepSeek V3/R1, Kimi (Moonshot), GLM (Zhipu). See [[Chinese Frontier Labs]].