private-local.ai vs LocalAI: free software or a tested machine
Benjamin Spinola, Founder
Published Updated 9 minutes read
private-local.ai vs LocalAI appears to be a choice between a paid service and free software, but the two answer different questions. LocalAI is open-source software under a permissive licence that turns a machine you already own into a server that answers requests to artificial intelligence (AI) models. What it can do is limited by that machine, and the computers that most companies already own are usually too small for a capable open model.
LocalAI runs models. It does not tell you which machine to buy, how large it needs to be or whether your particular work performs well on it at your real volume. Those are the expensive questions, and they are the ones we answer. We operate an NVIDIA DGX Spark with a named open model and provide a standard endpoint by the hour, so that you can connect your own tools and run your own tasks on a configuration that you could reproduce on equipment that you buy.
All information about LocalAI on this page comes from the project's own website and repository, read on 22 September 2026.
private-local.ai vs LocalAI: the main differences
| Criterion | private-local.ai | LocalAI |
|---|---|---|
| What is provided | You receive temporary access to a named machine, model and serving configuration through a web address and an access key. The configuration is documented, so that a result can be reproduced later on equipment that you own. | LocalAI is a server for running models, known as an inference server, which you download and run under a permissive open-source licence. It places one programming interface in front of more than 60 interchangeable engines and offers a gallery of models that can be installed with one click. |
| Who operates the system | We operate the machine and the model for the duration of the test drive, and you connect your own tools and run your own workflows. If you then buy a machine, your company operates it. | Your organisation operates it entirely on its own. The project publishes no support tier and no compliance certification; it has two named maintainers, 237 code contributors and two corporate sponsors. |
| Where the model runs | During the test drive, the model runs on the machine that you selected, which we operate on your behalf. | The model runs on whichever device you install LocalAI on, whether a laptop, a server, a single-board computer or a cluster. By default, LocalAI sends nothing to any external service, so no data has to leave the building. |
| Which models are used | The test drive uses open models that we name and fix to a specific version, selected to fit the hardware being tested, so that a result can be attributed to an exact configuration. | LocalAI runs open models, and its gallery listed 1,255 of them on 22 September 2026. The project also builds its own compressed versions, so that larger models fit on smaller graphics cards, and publishes those files openly. |
| How the costs are structured | The test drive is charged at an hourly rate, which is published on our pricing page. What you pay for is an answer to the purchase question rather than computing capacity. | The software is free. The costs are the hardware, the electricity and the time of an engineer, and the last of these is usually the largest and the least visible. |
| Which organisations it suits | It suits a company that is considering a hardware purchase and wants to see the exact machine perform its work first, without anyone in the company having to install and serve a model on it. | It suits teams that have an engineer, a machine powerful enough for the model they need and the time to configure it. It is also a reasonable component of a production setup. |
What LocalAI is
LocalAI is an open-source inference server whose first code was committed in March 2023 and whose development is led by Ettore Di Giacinto. On 22 September 2026, its website showed version 4.8.0 under the permissive MIT licence, and its repository on GitHub showed about 49,000 stars, a common measure of interest, and 237 people who had contributed code. Its chat community had about 3,200 members. Its headline is that AI should run on every machine.
The design explains why the project covers so much ground. A small core accepts the same request format as the major cloud providers, so an existing tool usually needs only its address changed. Behind that single address sits a choice of more than 60 engines, which are loaded only when a model requires one. The same request can therefore be served by one engine today and by a different one tomorrow without any change to the tool.
The scope extends well beyond chat. It includes live transcription with speaker labels, speech synthesis and voice cloning, object detection and depth estimation from ordinary photographs, image and video generation, an agent built into the program that can call tools and load plugins, and a filter that removes names, addresses and card numbers on the machine before anything is sent elsewhere. The range of hardware is equally broad, from a single-board computer to a cluster, and since April 2026 several machines can be pooled, so that each request is processed wherever the model is already loaded.
One detail matters to anyone considering a DGX Spark. The project builds its own compressed versions of models and publishes benchmark tables for them, measured on an NVIDIA DGX Spark, the machine on which our test drive runs. In its published run, the model Qwen3.5-35B-A3B, with 35 billion parameters, shrinks from 64.6 to 12.2 gigabytes while its speed more than doubles, from 30.4 to 74.4 tokens per second, and its quality scores change only slightly; a token is a unit of text of roughly three quarters of an English word. These are the project's own measurements on its own machine, and they show how much the configuration matters on identical hardware.
LocalAI's strengths
- LocalAI is free, mature and broad: for a technically confident team with a capable machine, it covers everything needed to run models on the team's own hardware.
- Running on the processor alone is treated as a fully supported case rather than a degraded fallback and is tested in every build, so trying LocalAI does not require buying a graphics card first.
- The project publishes open benchmark results, together with the method, measured on the same machine that our test drive uses.
LocalAI's limitations
- LocalAI is software, not a supplier. It has no published support tier and no compliance certification, and nobody is accountable if a model gives a wrong answer in a critical situation.
- It does not tell you which machine to buy, how large it needs to be or whether your particular work performs well on it at your real volume, and those are the questions on which a purchase depends.
- Its breadth is also a burden. More than 60 engines and more than 1,000 models form a large space of possible configurations, and somebody has to find the right one for a specific workload before the results mean anything.
Which option suits which organisation
LocalAI is the better choice if
- You have an engineer who is willing to do this kind of work and a machine powerful enough for the model you need.
- You only need small models that run well on the computers you already have.
- Your question is whether a model can do the work at all, rather than whether a specific machine is worth buying.
Our test drive is the better choice if
- You have already established that the software works, and the remaining questions are which machine to buy, at what size, and whether the result holds at your real volume.
- Nobody in your company has the time to build a test setup, and the hardware that you would be testing costs several thousand dollars to buy on an assumption.
- You need a written record of what ran, on which machine and how well, in a form that the person who approves purchases will accept. Free software does not produce such a record.
Sources and notes
All information about LocalAI on this page was read from the project's own website and public repository on 22 September 2026, including the version number, the licence, the community figures, the list of engines and the published benchmark table. These are the project's own statements and measurements, and they are reported as such.
LocalAI is open-source software that we may use as a component of the configurations that we prepare. We have no commercial relationship with the project.
Our own measurements will be published together with the machine, the model version, the serving software, the task and the date. If any information on this page is out of date, please let us know and we will correct it.
Further reading: the other four comparisons cover ChatGPT, Palantir, the redaction company Limina and the on-premises assistant Seesa, and the pricing page states the hourly rate.
Questions about private-local.ai and LocalAI
private-local.ai vs LocalAI: if the software is free, what does private-local.ai charge for?
Not for the software. LocalAI runs models, but it does not tell you which machine to buy, how large it needs to be or whether your work performs well on it at your real volume. We operate that machine for you by the hour with a specific, disclosed configuration, and you connect your own tools, run your own tasks on it and keep a written record of what ran, on which machine and how well. A purchase decision needs that record, and free software does not produce it.
Could we install LocalAI ourselves instead?
Yes, if you have a capable engineer and a machine powerful enough for the model you need. Most office computers are not powerful enough, because a capable open model needs far more memory than they have. If you are unsure which machine would be sufficient, our test drive answers that question with a prepared machine, a served model and a working endpoint, so that you can measure before you buy.
Do you use LocalAI yourselves?
We may use it as a component of the configurations that we prepare, and we would name it explicitly if we did, because the purpose of the test drive is that you can reproduce the result on your own equipment. We have no commercial relationship with the project and have not sponsored it.
Find out which machine suits your workload
If you already run LocalAI, the software question is settled. Choose the setup closest to the machine that you are considering and run a representative task on it; most questions are answered within one or two hours.
Request an endpoint