Open-weight models explained

0:00 Open-weight models

Some of the most interesting AI in twenty twenty six is not behind a subscription at all. You can download it, run it on your own servers, and customise it. In this lecture you will learn what open weight models are, when they beat closed models for a business, the licence traps to avoid, and how model size and hardware fit together.

0:27 Why this matters

Why does this matter? Three reasons come up again and again. Some data cannot leave your control, because of client contracts, regulation or sheer sensitivity. Some tasks run at such high volume that per token pricing adds up. And depending on a single vendor means their price or policy change becomes your problem. Open weight models give you another option for each of those situations.

0:55 Three kinds of model

Let's get the vocabulary right. Closed models are used through a vendor's app or API, and you never see the weights. Open weight models publish their trained weights for download under a licence. Examples include Meta's Llama, Mistral, Alibaba's Qwen, DeepSeek, Google's Gemma and OpenAI's g p t oss models. And open source in the strict sense, as defined by the Open Source Initiative, also requires enough information about training data and code to study and modify the system. Many open models are open weight, but not fully open source.

1:34 The analogy

Here is an analogy. A closed model is like eating at a great restaurant. Excellent food, no kitchen to run, but you eat what is on the menu and you pay per plate. An open weight model is like getting the recipe. You can cook it at home, adjust the spices and cook in bulk more cheaply. But you need a kitchen, you need to learn to cook it well, and you do the washing up. Neither is better in general. It depends on what you are serving and how often.

2:14 Benefits

The benefits are real. Control and privacy, because data can stay on your servers or in your chosen cloud region. Customisation, because you can fine tune on your own data and style. Cost at scale, because for high volume, narrow tasks, self hosting can be cheaper than per token pricing, depending on how busy your hardware is. Independence from a single vendor. And even offline or air gapped use for sensitive environments.

2:45 Trade-offs

And the trade offs. The strongest closed frontier models usually still lead on the hardest reasoning and agent tasks, although the gap varies and narrows with each release. You run the operations, hosting, scaling, monitoring and security patches. You own safety, the guardrails and filtering a vendor would otherwise provide. And you must follow the licence, which brings us to the part people skip.

3:13 Read the licence

Licences range from permissive, like Apache two point zero or MIT, used by several Mistral, Qwen, DeepSeek and g p t oss models, to custom licences with conditions, such as Meta's Llama community licence and Google's Gemma terms. These may include acceptable use policies, attribution requirements, or restrictions for very large services. Before commercial use, find the official licence, check commercial use, attribution, acceptable use and any user count limits, check whether fine tuned versions inherit it, and record the licence version with the model you deploy.

3:51 Simple example

A simple example. A writer wants to summarise her private journal entries into weekly themes, but she does not want them on anyone's servers. A small open weight model running on her laptop, offline, does the job well. Summaries of personal notes do not need a frontier model. She checks the licence, which is permissive, and she is done. Small model, right job, full privacy.

4:19 Sizes and hardware

Now sizes and hardware. Models are measured in parameters, from a few billion to hundreds of billions. Bigger is usually more capable and needs more memory. Mixture of experts models activate only part of the network for each token, which helps speed. Quantisation compresses the weights, for example to eight or four bit, so a model fits in less memory, with some quality loss. The rough rule is that the model must fit in your GPU memory, or unified memory on Apple Silicon, with room left for the context.

4:58 Business example: a London law firm

A realistic business example. A mid size law firm in London needs to classify thousands of internal documents, but its client contracts forbid sending them to external AI providers. It deploys an open weight model on its own cloud tenancy and fine tunes a small model on two thousand labelled examples. Tested on three hundred held out documents, the small model meets the accuracy bar for tagging. Complex legal analysis still goes to lawyers. The licence and data flow are documented for the risk committee. A year later the firm reviewed the decision. A newer open weight model had been released under the same permissive licence, and the firm reran its three hundred document evaluation. The newer model scored higher, so they upgraded, updated the register with the new version and licence, and repeated the risk committee sign off. The process made the upgrade routine rather than risky.

6:02 Common mistakes

Four common mistakes. Assuming open means no conditions. Choosing the biggest model your hardware can barely run, which is slow and fragile. Skipping evaluation because it is our own model. And forgetting guardrails and security on self hosted endpoints, which then sit on the internet with no protection at all.

6:24 Watch me do it, part 1

Let me check a licence the way the law firm did. I open the model's page on Hugging Face. The model card shows the parameter count, the available formats and quantised versions, and a licence field. I click the licence and read it. Commercial use, permitted. Attribution, required in the documentation. Acceptable use policy, a list of prohibited uses. And for fine tuned versions, the same licence applies. I record the model name, version, licence and date checked in our register, alongside the risk committee's approval.

7:01 Watch me do it, part 2

Next, size. Our server's GPU has limited memory. The eight bit version of the larger model does not fit with room for context, so I mark it red. The four bit version fits, and a smaller model fits easily. Then the evaluation plan. Three hundred held out documents, already labelled by paralegals, and an accuracy target agreed with the practice lead. After fine tuning, the small model meets the target on tagging. Our notes say clearly, suitable for document tagging, not for legal analysis, which stays with lawyers.

7:39 Recap and try this now

Recap. Know the difference between closed, open weight and open source. Choose open weight for control, customisation, cost at scale and independence, and accept the operational and safety work. Read the licence every time, and size the model to your hardware. Try this now. Pick one open weight model family, find its official licence page, and summarise in three bullets whether and how it permits commercial use, and any conditions you would need to follow.