An open-weight model is an AI model whose trained weights, the learned numbers that make it work, are publicly released so anyone can download, run and often fine-tune it. Examples include model families such as Llama, Mistral, Qwen, Gemma and DeepSeek. Open-weight is not always the same as open source: many releases share weights but not training data or full training code, and licences vary widely.
Key takeaways
- Open weights let you run models on your own hardware or cloud.
- “Open-weight” and “open-source” are often confused but differ.
- Licences range from permissive (such as Apache 2.0 or MIT) to custom licences with usage restrictions.
- Benefits include privacy, control, customisation and cost; risks include misuse and security responsibility.
- Closed models are accessed only through an API or app.
An open-weight model explained in simple terms
A closed model is like a restaurant: you can order dishes, but you never see the recipe or the kitchen. An open-weight model is like receiving the finished recipe card: you can cook the dish at home, adjust it and serve it however you like, subject to the licence. A fully open-source model goes further, sharing how the recipe was developed, including the ingredients list and the kitchen methods.
Open-weight vs open-source vs closed
| Aspect | Closed model | Open-weight model | Fully open-source model |
|---|---|---|---|
| Weights downloadable | No | Yes | Yes |
| Training code | No | Sometimes | Yes |
| Training data details | No | Usually limited | Yes, or detailed description |
| Licence | API terms | Varies, sometimes restrictive | OSI-approved style freedoms |
| Where it runs | Provider’s servers | Anywhere you choose | Anywhere you choose |
In October 2024, the Open Source Initiative published version 1.0 of its Open Source AI Definition, which requires access to model weights, code and sufficiently detailed information about training data. Many “open” models do not meet this bar, which is why “open-weight” is the more accurate term for them.
A short history
- 2023: Meta released Llama 2 with downloadable weights under a custom licence, and Mistral AI released Mistral 7B under the permissive Apache 2.0 licence.
- 2024: Open-weight releases accelerated from Meta, Mistral, Google (Gemma), Alibaba (Qwen) and others, narrowing the gap with closed models.
- 2025: DeepSeek released its R1 reasoning model with open weights under the MIT licence, drawing global attention to open-weight reasoning models.
Why organisations use open-weight models
- Data privacy: run models on-premises or in a private cloud so data never leaves your control.
- Customisation: fine-tune with methods such as LoRA on your own data.
- Cost control: at high volume, self-hosting can be cheaper than per-token API fees.
- No lock-in: avoid dependence on one provider’s pricing and model changes.
- Offline and edge use: run quantized models on laptops or devices.
- Transparency and research: inspect and study model behaviour.
A worked example: choosing an approach
A hospital group wants a model to summarise patient notes. Sending records to an external API raises data protection concerns. They select an open-weight model with a licence that permits commercial use, run it on their own GPU servers, fine-tune it on de-identified notes with LoRA, and quantize it to cut hardware costs. They also take on responsibility for security, updates, monitoring and evaluation.
Risks and responsibilities
- Misuse: once released, weights cannot be recalled, and safety guardrails can be removed by fine-tuning.
- Licence compliance: some licences restrict use by very large companies, certain use cases or require attribution.
- Operational burden: you manage hosting, scaling, inference optimisation and security.
- Quality gap: the best closed models may still outperform open ones on some tasks.
Limitations
- Hardware for large models is expensive.
- Training data is often undisclosed.
- Updates and patches depend on the releasing organisation.
- Safety is the deployer’s responsibility.