Glossary · AI & automation
Open-weight model meaning
An open-weight model is an AI model whose trained weights can be downloaded and run on your own servers. Learn how it differs from open source and hosted APIs.
What does open-weight actually mean?
A language or vision model is, at bottom, a very large file of numbers learned during training. Those numbers are the weights. With a hosted model, the weights stay with the provider and you send your text or images to its servers over an API. With an open-weight model, you download the weights and run inference on hardware you control, so the document never has to leave your network to be read.
Open-weight describes access to the weights only. It says nothing on its own about the licence, the training data or how good the model is at your task.
How is an open-weight model different from open-source AI and a hosted API?
| Type | What you get | Where it runs |
|---|---|---|
| Hosted model (API) | Answers, not weights | Provider's servers |
| Open-weight model | Weights under a licence that may restrict some uses | Your servers, VPC or a hosting partner you choose |
| Open-source AI (OSI definition) | Weights plus code and enough information about the training data to rebuild and modify it | Anywhere the licence allows |
Many models described as "open" are open-weight but do not meet the Open Source Initiative's definition, because the training data details are withheld or the licence limits commercial use.
Why do banks and NBFCs in India ask for open-weight models?
When a model reads a PAN card image or a bank statement, the security team asks where inference runs, what the model receives, and whether anything is retained or used for training. Some institutions accept a hosted model with contractual limits; others require a local or open-weight model inside their own environment. The on-premise vs cloud AI guide suggests writing this as a policy by data class rather than one rule for everything. Products such as digital KYC on WhatsApp from AutoKYC can run on managed cloud, your own VPC or on-premises, with local or open-weight models when policy demands it.
How much hardware does an open-weight model need? A worked example
Illustrative arithmetic, not a sizing recommendation. Memory for the weights is roughly the number of parameters multiplied by the bytes stored per parameter. A model with 8 billion parameters held at 16-bit precision uses 2 bytes each: 8 billion × 2 = about 16 GB. Compressed to 4-bit (half a byte each), the same model needs about 4 GB for weights. Real deployments need extra memory for the text being processed and for several requests at once, and must be sized for month-end peaks, not the average day. Compression saves memory but can reduce accuracy, so test on your own documents.
What do teams get wrong about open-weight models?
- Assuming downloadable means free for any use. Read the licence. Some restrict commercial use, user counts or certain sectors.
- Assuming local means safe. A model on your own server still needs access control, logs, key management and a read-only default for actions, as the read-only KYC agent guide sets out.
- Skipping evaluation. A smaller self-hosted model may misread a blurred document that a larger hosted one reads correctly. Measure on your own files.
- Forgetting maintenance. Updates, security patches and re-testing become your job, not the provider's.
- Training on customer data casually. Fine-tuning with personal data raises consent and purpose questions under the DPDP Act. Confirm with your legal adviser.
Related terms
Read next
Where this shows up in our work. AutoKYC: Documents read on WhatsApp, checked against your policy, exceptions to a person, maker-checker before the core. Book a 30-minute working session with an engineer.
Questions
What is the difference between open-weight and open-source AI?
An open-weight model publishes its trained weights so you can run it yourself. Open-source AI, as defined by the Open Source Initiative, also requires the code and enough information about the training data to study, rebuild and modify the system.
Can a bank run an open-weight model on its own servers?
Yes, that is the main reason to choose one. The bank downloads the weights and runs inference inside its own data centre or private cloud, subject to the licence terms and its own security and data policies.
Are open-weight models less accurate than hosted models?
Not always. Accuracy depends on the specific model, its size and the task. Smaller or compressed models can trail larger hosted ones on hard documents, so test both on a sample of your own files before deciding.