Open Weights
also called Open-source model
A model whose trained parameters you can download and run yourself.
A model published with its weights available, so anyone can download, run, modify, and often fine-tune it. Llama, Mistral, Qwen and DeepSeek are common examples.
Open weights means you control data residency, cost and customisation, at the price of running your own hardware and taking on the operational burden.
Be precise about which freedoms you are getting. Published weights give you the ability to run, inspect and fine-tune — that is real and valuable — but they do not automatically give you training data, the training recipe, or the right to do anything with the output. Those are separate grants, and the licences differ per model.
Choosing open weights for a reason
A regulated team cannot send customer records to a third party.
Weights you can run in your own network remove that objection entirely, and fine-tuning adapters can be kept alongside the data. You buy back control and pay for it in ops: GPUs, quantisation choices, upgrades you now own.
Calling it "open source" without qualification — the licence may restrict commercial use or the training data is not released.
- paperLlama 2: open foundation and fine-tuned chat models ↗
A released family from 7B to 70B plus chat variants, with the fine-tuning and safety work documented. A primary example of published weights — not a ruling on what "open source" should mean.
- paperMistral 7B ↗
Grouped-query attention and sliding-window attention in a single model write-up: the concrete engineering behind cheaper inference and bounded attention.
- paperQLoRA: efficient finetuning of quantized LLMs ↗
Backpropagate through a frozen 4-bit quantised base model into LoRA adapters, which is what makes single-GPU fine-tuning of large models ordinary.
Foundation Model · Base model
A large model trained broadly, meant to be adapted to many downstream tasks.
Fine-tuning
Extra training on your own examples to change how the model behaves.
Quantization
Storing weights at lower precision to make a model cheaper to run.
Vendor Lock-in
How hard it would be to leave your model provider.