Open-weight vs closed AI models: what changes when you train on your own data

A closed model (OpenAI's GPT models, Anthropic's Claude, Google's Gemini, Amazon Nova) is reached only through its provider's service. You never see the weights. If the provider offers fine-tuning, you send your examples, the provider trains a private copy, and you get back a model name you can call — on their platform, at their prices, for as long as they keep it.

An open-weight model (Qwen, Llama, Mistral, Gemma, gpt-oss and others) publishes its weights for anyone to download, under a license. You, or a service acting for you, can train it on your examples — usually as a small LoRA adapter on top of the original weights — and keep the result. Open weights aren't always open source: the training data and code may not be published, and licenses differ, from permissive Apache 2.0 to community licenses with conditions. Read the license before you build on one.

On May 7, 2026, OpenAI began winding down self-serve fine-tuning. According to its API deprecations page, checked October 26, 2026:

• May 7, 2026: organizations that had never fine-tuned can no longer start.
• July 2, 2026: organizations without recent inference on a fine-tuned model can no longer start new jobs.
• October 23, 2026: several older fine-tuned models were shut down.
• January 6, 2027: active customers can no longer create new fine-tuning jobs.
• January 14, 2027: GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini are removed from the API — including their fine-tuned variants on the same retirement dates.

None of this is unusual or unreasonable: providers move to newer models, and old ones are expensive to keep serving. But it shows the trade plainly. A closed fine-tune is rented. Its life is the life of its base model, on a schedule you don't set. (Microsoft runs its own schedule for the same models on Azure; check Microsoft’s retirement page for dates there.)

An open-weight adapter has a different failure mode. It doesn't get switched off, but it is tied to one exact set of base weights: load it onto a different revision of "the same" model and it can still return confident answers that are quietly wrong. Pinning the exact base, and keeping a copy of it, is part of owning one.

Here is the part that matters most, whichever kind you choose. Your labeled examples and held-out test are the durable assets. The spreadsheet of past decisions your team made, the corrections your reviewers add, and a held-out test set that tells you honestly how a model does on rows it never saw.

With those, any model is replaceable: when a base model retires, or a better one arrives, you train again on the same examples, score the new model on the same held-out rows, and switch only if it's better. Without them, every model change is a leap of faith.

So the useful questions are:

1. Where do my examples live, and can I retrain from them in an afternoon?
2. Is every model tested the same way, on rows it never trained on?
3. Can I try a closed model and an open-weight one on the same data and compare?

StayCharted AI Model Trainer starts from exactly that asset: your team's past decisions, in a spreadsheet or a folder of photos. It cleans them, holds out test rows, and lets you build several kinds of model from the same examples — some on open weights, some on closed ones — each scored on the same held-out rows, so you can choose by evidence.

Can it help? Yes, in four ways

1. One set of examples, every kind of model. Upload once. Train an open-weight model and a closed one, or ask your own closed model with your examples, and compare their reports on the same held-out rows.
2. Open weights without running GPUs. The Dedicated AI Model fine-tunes an open-weight base on StayCharted's GPU fleet. You get the control of open weights — a pinned base, an adapter you can download — without hiring for it.
3. Closed models, kept in your account. If your company has standardized on Bedrock, Foundry or Google Cloud, StayCharted fine-tunes there: the training file goes to your account and is deleted when training ends, and the tuned model never leaves it.
4. Ready when a base model retires. Your examples, your reviewers' corrections from the Review Queue, and your test rows stay in StayCharted. When a base model is retired — open or closed — you retrain on the same data, compare the new version with the live one, and publish it only if it's better.

Where it doesn't help

• Open weights don't make answers free. Someone runs the GPU. On StayCharted the fleet is shared and the cost is in your plan; running a downloaded adapter yourself means your own GPU and serving stack, and StayCharted does not provide support for serving downloaded models yourself. A self-hosted Enterprise deployment of the whole product is a separate offering; contact us to discuss it.
• Closed models can know more. For categories that depend on general world knowledge, a large closed model asked with good examples may beat a small fine-tuned one. That's why Your AI Models exist, and why every kind gets the same test.
• StayCharted doesn't fine-tune OpenAI or Claude directly. It uses your own key to ask them, with your examples in each question.

• Your team's own conventions, lots of examples, data that must stay private: an open-weight fine-tune (Dedicated AI Model), or an AI Classifier Model for short everyday text.
• Company standardized on one cloud, and the model must live there: a Dedicated AI Model trained in your own Bedrock, Foundry or Google Cloud.
• Categories that lean on general knowledge, few examples yet: Your AI Model, with your past decisions as examples.
• Not sure: build two from the same file and let the held-out results decide.

Start free → · See the models · Train in your own cloud · Use your own AI