Definition
Open-weight models are AI models whose trained parameters, the weights, are released publicly so that anyone can download them, run them on their own hardware, inspect them and fine-tune them. They contrast with proprietary models such as the frontier versions of Claude, GPT or Gemini, which you can only use through the provider's API. The term is more precise than open source, because the training data and code are often not fully released.
For a company, open-weight models mean control. You can run a model inside your own cloud or data centre so that sensitive data never leaves, fix a version so behaviour does not change under you, fine-tune it for a narrow task, and avoid per-token pricing at high volumes. The trade-off is that you take on hosting, GPUs, security and updates yourself, or pay a provider to host the open model for you.
In 2026 open-weight models from Meta (Llama), Mistral, Alibaba (Qwen), DeepSeek, Moonshot (Kimi) and others sit close behind the proprietary frontier and ahead of it on cost for many tasks. The gap is small for routine work and larger for the hardest reasoning tasks. The misconception is that open-weight means free. Weights cost nothing, but running them well costs engineering time and hardware.
In practice
A European bank runs an open-weight model on its own servers to summarise customer conversations, satisfying regulators that no data leaves its infrastructure, while using a proprietary frontier model through an API for tasks that involve no personal data.
Why it matters
Open-weight models give you a negotiating position and an exit from any single AI vendor. For regulated or high-volume use cases, they are often the sensible default.
Frequently asked questions
- What is the difference between open-weight and open-source AI?
- Open-weight means the trained model parameters are published and can be run and modified. Fully open-source AI would also release the training data and code under an open licence. Most popular open models are open-weight with a licence that may restrict some commercial uses, so read it.
- Are open-weight models as good as ChatGPT or Claude?
- For many everyday business tasks, the best open-weight models are close and far cheaper to run at scale. The proprietary frontier models still lead on the hardest reasoning, coding and agent tasks. Many companies use both, routing each task to the model that fits.