Open-Source vs Hosted AI: How to Choose a Model

Quick answer: choose a hosted model when you want the least setup and your data is allowed to leave your systems. Choose an open-weight model when you need control, offline use, data residency or predictable cost at steady volume, and you can handle the extra work. Many teams start hosted and move individual workloads to open models later. Whichever way you lean, read the licence: "open" covers very different terms.

Data checked: 7 November 2026. Counts come from the What's That AI model catalog (4,400 models). Open-source flags are imported from source records, and licence and download figures come from Hugging Face records collected between 17 and 27 September 2026. This is a guide to choosing, not legal advice. See how our data works.

Hosted, open-weight and open source are not the same thing

People use these terms interchangeably, and that is where bad decisions start.

• Hosted model. You reach it through an app or API. The provider runs it, and you never receive the weights.
• Open-weight model. You can download the weights and run them yourself. The licence may still limit commercial use or add acceptable-use rules.
• Open-source model. The Open Source Initiative's Open Source AI Definition asks for more than downloadable weights, including enough information about the training data and the code. Many models labelled "open" meet only the open-weight part, so read the licence rather than trusting the label.

What the catalog shows

Of 4,400 models in the catalog, 3,545 (80.6%) carry an open-source flag, 800 (18.2%) do not, and 55 (1.3%) have no flag. That is not a picture of the whole market. 3,424 of our models were matched to Hugging Face records, which skews the catalog toward open releases. Read the figures as a view of what you can find and compare here.

Open models are not equally common in every kind of task.

Text and audio are dominated by open entries. Video is split roughly in half, and coding is a small category of 40 entries, so read that row cautiously. You can see the split yourself with the open-weights filter and the closed filter, or narrow by type, such as audio or video.

Licences are where "open" gets complicated

2,962 catalog models have a licence recorded. Grouped by family:

About 22% of licensed models fall outside the permissive group, so a quick licence check is not optional. Examples from the catalog: Qwen3-0.6B is recorded as Apache 2.0, GPT-2 as MIT, Stable Diffusion XL under OpenRAIL++, and Gemma 3 1B under the Gemma terms. The licence on the model page is a starting point; confirm it at the source before shipping.

Downloads show where open models quietly do the work

Look at the 20 most-downloaded models in the catalog and you will find mostly small specialists: sentence embedding models, rerankers, encoders such as BERT, and compact classifiers. The most-downloaded entry, all-MiniLM-L6-v2, shows about 244 million downloads in our records. Another, BGE M3, is an embedding model under an MIT licence. They power search, matching and tagging inside products rather than chat windows.

The lesson for buyers is that your job may not need a large chat model at all. A small open model for a defined task is often cheap to run, and download counts are a usage signal rather than a quality ranking.

Hosted or open: the practical trade-offs

There is a middle path. Many hosting providers serve open-weight models by the request, which gives you the licence flexibility without owning a server. It does not remove the data question, because your inputs still leave your network.

Will it run on your hardware?

A rough rule: memory for the weights is the number of parameters multiplied by the bytes stored per parameter. Sixteen-bit weights take 2 bytes each, 8-bit takes 1 and 4-bit takes about half a byte.

These are weights only. Context length, the runtime and your operating system add overhead, so leave headroom. Quantising to fewer bits can also lower quality, so test the compressed version on your own prompts before you rely on it.

A six-step way to choose

• Define the job. Chat, search, transcription, classification, image generation? The catalog's text-generation filter is only one of many task types.
• Settle the data rule. If inputs cannot leave your network, you are looking at self-hosted open models or a hosted service with terms you have reviewed.
• Read the licence. Check commercial use, user or revenue thresholds, and acceptable-use rules.
• Try before you download. The demos catalog lists 790 runnable demos, so you can test an open model on your own sample before setting anything up.
• Cost it at your volume. Estimate monthly requests, then compare per-use pricing with hardware or hosting plus your time. At low volume, hosted is usually cheaper. At steady high volume, self-hosting can win.
• Run a 20-prompt bake-off. Use real inputs, hide the model names, and have two people rate the outputs. Keep that prompt set so you can re-test when a new model arrives.

Frequently asked questions

Is open-source AI free?

The weights are often free to download. Running them is not free: you pay for hardware, electricity or a host, plus the time to maintain it. The licence may also impose conditions.

Are open-weight and open source the same?

No. Open-weight means you can download the model. Open source, as the Open Source Initiative defines it for AI, asks for more, including information about training data and code. Check which one a model actually offers.

Are open models as good as hosted ones?

It depends on the task. For narrow, well-defined jobs, open models can be excellent and inexpensive. For the hardest open-ended work, hosted models are often ahead, but rankings change quickly. Test on your own data.

Can I run a model on a laptop?

Small models, especially quantised ones in the 1 to 8 billion parameter range, can run on recent laptops. Use the memory table above to estimate, and expect slower responses than a hosted service.

Is self-hosting more private?

It keeps inputs under your control, but security is then your responsibility. Patching, access control and logging become your job.

What does the "closed" filter on the models page mean?

It selects entries whose imported source flag says they are not open source, which usually means a hosted or proprietary model. The flag comes from source data, and 183 closed-flagged entries still have a Hugging Face record, so confirm details on the model page.

Your next step

Start with the job, not the label. Browse open models or the full model catalog, shortlist two candidates, and run the bake-off above. If you would rather use a ready-made product, compare tools by task in the AI tool directory.

Budget matters for either route. Our guide to free AI tools and what free really means explains how to test a free plan before you pay.