Open weights
In one sentence An open-weights model is one whose trained parameters are published for download, so anyone can run and fine-tune it — which is not the same as open source.
Updated
An open-weights model is one whose trained parameters — the weights file — is published for anyone to download, run and adapt.
A useful analogy: a restaurant can hand you the finished dish, the recipe, or both. Closed API models (GPT-4-class, Claude) serve the dish — you consume the capability over the network and never touch what is underneath. Open-weights models (Llama, Mistral, Qwen, DeepSeek, Gemma) hand over the cooked result itself: the billions of learned numbers. You can run it on your own hardware, fine-tune it on private data, quantize it for a laptop, and never send a customer's data to anyone.
The pedantic distinction in the name is real: open weights is not open source. The recipe — training data, full training code — usually stays private, and many releases carry licences with conditions (Llama's community licence, research-only clauses). Truly open-source releases (weights, data recipe, code — e.g. OLMo) are rarer. Read the licence before shipping a product.
Why the ecosystem matters practically: privacy and data-residency (medicine, banking, government), cost control at scale, offline and edge deployment, and the ability to modify behaviour deeply. The trade-offs: you own the serving problem — GPUs, throughput, updates — and the strongest frontier capability tends to reach closed APIs first, with open models following months behind. The distribution rails are worth knowing by name: Hugging Face hosts the checkpoints; GGUF files and Ollama put them on laptops.
Where to go next
- Full lesson: Ollama
- Related terms: checkpoint, foundation-model, fine-tuning, gguf