TL;DR: An open-weights model publishes its trained parameters for anyone to download, run, and fine-tune. That is not the same as open source - the training data and code usually stay private, and licenses vary - but it is enough to run AI on your own hardware, own your stack, and never send data to a third party.
How it works
A model's weights are the billions of numbers training produced - the entire learned capability in one artifact. Closed providers keep theirs behind an API: you send a request, they run the model, you get tokens back. An open-weights release ships the artifact itself, typically as safetensors files on Hugging Face, and from there the model is yours to operate: serve it with vLLM on rented GPUs, run a quantized build on a laptop through Ollama, or fine-tune it on data you could never ship to an outside API. Major families in this world include Llama, Qwen, DeepSeek, Mistral, and OpenAI's gpt-oss line.
The phrase "open weights" is deliberately narrower than "open source", and the distinction has teeth. True open source would include the training code, data recipe, and an OSI-approved license; almost no competitive model clears that bar. Open-weights releases give you the finished parameters under licenses that range from genuinely permissive (Apache 2.0 - Qwen and Mistral's recent releases, DeepSeek's models) to custom terms with usage restrictions and commercial thresholds (Llama's community license). You can run and modify the model, but you cannot reproduce it, fully audit what went into it, or always use it however you like - read the license before building a business on one.
Why choose open weights when frontier APIs are stronger and simpler? Four durable reasons. Data control: regulated and confidential workloads can run fully on-premises. Cost structure: per-token API pricing becomes fixed infrastructure cost, which flips the math at high, steady volume. Customization: deep fine-tuning and modification that hosted APIs do not allow. Permanence: a downloaded model cannot be deprecated, repriced, or silently changed underneath you. The trade is operational: you own serving, scaling, and safety, and the strongest open models have historically trailed the frontier - though the gap keeps narrowing and top open releases now land within striking distance of closed flagships on many benchmarks.
In practice many teams run a hybrid: a frontier API where peak capability matters, open-weights models for high-volume, private, or specialized paths - another axis in choosing a model.
Where it sits in the AI stack
Open weights change who operates the model layer - the artifact flows to your infrastructure instead of staying behind a vendor's API:
Key tools and implementations
-
Hugging Face Hub
The distribution point - model cards, weights, licenses, and community quantizations.
-
Llama, Qwen, DeepSeek, Mistral
The most-used open-weights families, spanning permissive and custom licenses.
-
vLLM and llama.cpp
The serving engines for datacenter-scale and laptop-scale self-hosting respectively.
-
Hosted open-model providers
Inference services that run open-weights models for you, per token, when you want the models without the ops.
Related entries
- Llama Meta's open-weight model family that developers can download, run locally, and fine-tune under a source-available license.
- Local LLM Running a language model entirely on hardware you control instead of calling a hosted API.
- Fine-tuning Further training a pretrained model on your own examples to specialize its behavior for a task.
- Qwen Alibaba's family of open-weight models spanning many sizes, strong in multilingual and coding tasks.