Open-weight and regional assistants

The previous lessons covered the assistants most people meet first, and they share two properties: they are hosted services, and they are made by large American companies.

There is a third category that most introductory courses skip entirely, and it matters most to exactly the people those courses serve least.

Open-weight models

A hosted assistant keeps the model on the company’s servers. You send text, you get text back, and you never have the model itself.

An open-weight model publishes the trained model — you can download it and run it on your own hardware. Several labs publish these, including Meta (Llama), DeepSeek, Mistral and Alibaba (Qwen).

Three consequences, and they are the reasons to care:

Nothing has to leave your device. E11·L5 made the cost argument and E12·L2 the privacy one. For confidential material, or where policy forbids external services, this is frequently the only compliant option.

No subscription and no per-token charge. You pay in hardware, electricity and setup effort instead.

Nobody can withdraw it. A downloaded model keeps working when a company changes its pricing, its terms, or its mind. E-P1·L8’s point about the ground moving under your prompt does not apply the same way.

The trade is capability: the best open-weight models sit somewhat behind the best hosted ones on hard reasoning, and comfortably ahead of what most everyday tasks require.

Open weights are not open source

Worth stating precisely, because the terms get used loosely and it has practical consequences.

Open weights means the trained parameters are published. Open source, in its usual sense, would also mean the training data and full freedom to use it.

Most of these releases publish weights under a custom licence with conditions — restrictions on commercial scale, on using outputs to train competing models, or on particular applications. The training data is almost never released.

So: free to download, not free of terms. Read the licence before building a business on one, and do not assume it works like an MIT-licensed library.

Regional providers

The other omission, and it connects two later modules.

E10·L1 noted that region-specific models frequently outperform general Western ones on their target languages. E11·L6 noted that local providers often price in local currency and accept local payment methods, which matters when an international card is declined.

Both point at a category the landscape lessons never showed you.

When a regional provider is worth checking:

Language. If you work mainly in a language the big services handle indifferently, test a provider built for it. DeepSeek and Qwen for Chinese, Mistral for European languages, and several national initiatives elsewhere.
Payment. Local currency, local methods, no foreign transaction fees.
Data residency. Where processing happens is a legal question in some sectors, and an EU-hosted or India-hosted option may be the one that satisfies it.
Cost. Some are markedly cheaper for comparable work.

What to check before relying on one: who operates it, where data is processed, what the terms say about training on your input, and how long it has been running. E7·L4’s source questions apply to vendors as much as to claims.

How to decide

Use E2·L2’s method — the same five tasks, run on a shortlist, judged on your own work. Add a hosted mainstream option, a regional or open-weight option, and compare honestly.

Most people never test the third category and assume it is worse. Sometimes it is. For a particular language, a particular budget, or a particular confidentiality requirement, it is frequently better — and you can only find that out by running your own tasks through it.

Try it now (7 minutes)

Take the five-task benchmark from L2. Run two of the tasks on an open-weight or regional option — the hosted chat interfaces are free to try.

Judge the result against your usual assistant. If the gap is small on the work you actually do, you have found either a cheaper or a more private way to do it.

Check your understanding

1. An open-weight model is one where: The source code of the company's website is public The trained model can be downloaded and run on your own hardware The training data has been published

2. Open weights and open source are: The same thing Different — weights are published, but the licence may restrict use and training data is usually withheld Different, but only in academic contexts

3. A regional provider is most worth checking when: You want the strongest possible reasoning Your language, payment methods or data residency requirements are poorly served You need the largest context window

Recap

Beyond the hosted Western assistants sits a tier most courses omit: open-weight models you can download and run yourself, and regional providers built for languages, currencies and jurisdictions the large services serve least well. Open weights are not open source — the licence has conditions. Test them with the same five tasks you used in L2, because assuming they are worse is how people miss a cheaper or more private option that would have suited them.