A private place for your models
Studio is a private place to use a model that runs on your own computer. You can chat, see which models you have, check which ones can look at pictures, and time how fast they answer.
It works with Ollama, LM Studio, llama.cpp, and other apps that run models on your machine.
The gap
Ollama, LM Studio, and llama.cpp can already run a model. Studio does not replace them. It does not install them or turn them on. The first time you open Studio, it prints the steps for your computer. You follow those steps yourself.
What is still left after the model app is installed:
Starting from zero
Each problem below is paired with what Studio does about it.
Ollama, LM Studio, and llama.cpp all run models. Each one installs differently and uses a different address. Picking one can feel permanent.
Studio The first run names those three. It prints the install and start steps for your computer only. It does not run those steps.
Guides jump to downloading a model before saying the app has to be running.
Studio It checks each app and says whether it is installed, running, stopped, or missing. Then it prints the next step.
A very large model will not fit on a laptop, and most guides never say what will.
Studio About 8 GB of memory is enough for a small model. About 16 GB is enough for a mid-size model plus a small one that can look at pictures. These are examples, not a model to go buy.
You cannot tell whether a chat left your computer.
Studio It stays on your computer. Chats, pictures, and speed tests are saved in a folder named data. There is no account, and nothing is sent back to us.
On Windows, typing localhost often opens the wrong address, so the first page looks broken.
Studio
Open http://127.0.0.1:8000. That address means this computer. Skip localhost.
Sending a picture to a model that only reads text fails with no explanation.
Studio It marks the models that can look at pictures. If a model cannot, the button to attach a picture stays hidden.
A first-run tool that installs software, or downloads a huge model, is a reason to close the window.
Studio It never installs or starts those apps. It downloads a model only when you ask, and only for Ollama.
After you install an app, each one has its own chat screen.
Studio Once any of those apps is running, Studio is the same place to work.
Already set up
Getting the model running was the first hurdle. Comparing models, keeping chats, and trusting the numbers are still hard.
Ollama, LM Studio, and llama.cpp do not share chats.
Studio One Studio install can talk to any of them. Older chats stay on your computer when you switch models.
A public ranking does not say how fast a model is on your computer, or how long you wait before the first words.
Studio It times one model at a time, on the app you saved. It leaves out a practice run by default. You can run the test once, or three times.
A quicker reply gets read as a better answer.
Studio The numbers are how fast this model answered on your computer. A finished run means a reply came back. It does not mean the writing or the picture is correct.
The model list is a set of names. It does not say which ones can look at pictures.
Studio You can download or remove a model only with Ollama. For the other apps, load the model there, then refresh. Studio marks which models can look at pictures.
A model can take a long time to download. Leaving the page looks like it stopped.
Studio A download keeps going in the background. Switching pages or refreshing shows the same progress. Only one download runs at a time.
At home or at work, the model may be on a different computer. Saving that address should be a choice you make on purpose.
Studio An address on this computer is saved right away. An address on another computer on your network is saved only after you confirm. The page says your messages will go there.
It is easy to paste a cloud address into a box that says “on this computer.”
Studio Saving the OpenAI website address is refused. The list you already saved stays as it was.
Chats buried inside an app are hard to back up or delete.
Studio Chats and pictures are files in a data folder on your computer. Deleting a chat deletes its pictures.
Different questions, a slow first try, or an older test make the table misleading.
Studio It compares two runs only when they used the same test, the same settings, and every question succeeded. Older tests can still be opened. They are not placed side by side.
Sample code should not run by itself, and a picture link should not be opened for you.
Studio Code is shown as text. It is never run. Pictures are made only by a picture app you already saved on your own network. Studio does not open picture links.
A sign-in stands between you and a model that is already on your computer.
Studio There is no sign-in, and Studio does not report how you use it.
What you get
Shows whether your model app is answering, how many models it has, and a download box for Ollama. If the app is stopped, the page stays up and shows the address to use. Start the app, then refresh.
The reply appears as it is written. The same chat is still there after you refresh. You can attach a picture only when the model can look at pictures. Deleting a chat deletes its pictures.
Lists the models on the app you saved, and marks which ones can look at pictures. Downloading and removing models is for Ollama only.
Times one model at a time. A practice run is left out by default. You can repeat the test once or three times. You see how long the first words take, and how fast the rest of the reply comes in. That is speed on your computer, not a score for how good the answer is.
Boundaries
Start
You need Python 3.11 or newer and Node.js 20 or newer. The install steps for Mac, Windows, and Linux are in the guide. Studio will not run them for you.
Then run:
python -m studio
Open http://127.0.0.1:8000 in your browser. That address means this computer. Skip localhost.
http://127.0.0.1:11434/v1
http://127.0.0.1:1234/v1
http://127.0.0.1:8080/v1
Any other app that runs models on your computer uses the address that app shows you. Keep the /v1 at the end of the address. Load the model in that app, then refresh Models.