Hey everyone, I built a UI client for Ollama to in...

# feed

[Kashismails]

Hey everyone, I built a UI client for Ollama to interact with deepseek models on your local machine. response handling, caching your chat history, model selection and all other basic functionalities are there already. Do try it out, contributions are welcome

https://github.com/Kashif-E/KOllama

[Stefan Oltmann]

Haha, right on time with the trend. 😄

[Kashismails]

didnt like the terminal interface, so thought of it as an opportunity 😛

[Meet]

yesterday i same idea thinging this project how to implement you created project

[Kashismails]

@Meet you can always contribute

[ade]

Very cool!! Out of curiosity, why does it run its own ollama process, rather than going via tcp like chatbox/open web ui?

[Robert Jaros]

https://github.com/Kashif-E/KOllama/issues/1

[Kashismails]

@ade because I didn’t think of it thats a nice idea and then i wont have to handle special characters as well pretty good idea ill look into it

[Kashismails]

Thanks @Robert Jaros for the issue will be fixing it shorty

[Meet]

@Kashismails remove this

state = rememberWindowState(width = 800.dp, height = 600.dp),

just window. minimumSize add it

Window(
        onCloseRequest = ::exitApplication, state = WindowState(
            placement = WindowPlacement.Maximized,
            position = WindowPosition(Alignment.Center)
        ), title = "Gemini AI KMP App",
        icon = org.jetbrains.compose.resources.painterResource(Res.drawable.gemini_logo)
    ) {
        window.minimumSize = Dimension(1280, 768)

        val mainViewModel =  koinInject<MainViewModel>()
        App(mainViewModel)
    }

[Kashismails]

thanks @Meet its fixed

[Kashismails]

@ade thanks for the idea its working much better now

[ade]

Sure! I have to say, that was incredibly quickly fixed 😄 good job

[Frank Bouwens]

Very nice 😄 !

[Kashismails]

Thanks @Frank Bouwens

[Kashismails]

@ade the idea was great and made me excited

[Stefan Oltmann]

I would actually like to try it and run deepseek locally.
My hardware is a PC with an RTX 2080 as well as an MacBook M1 8GB.

I know neither won’t run the biggest model of course, but hearing that one model even runs on RasPI I’m asking for advice which model might run on one of my devices.

[Frank Bouwens]

Just install ollama, and run ollama run deepseek-r1 by default it will install the 4,7gb version.

[Robert Jaros]

Today I've tested deepseek-r1:32b on my laptop (thinkpad with 32GB ram and Quadro T1000).

[Stefan Oltmann]

And will that work or is my hardware to weak?

[Frank Bouwens]

Works on my M2 MacBook

[Stefan Oltmann]

I guess the 8 GB won’t work or will they?

[Stefan Oltmann]

Just asking before I spent the whole evening and end up frustrated 😄

[Frank Bouwens]

Ehm my M2 MacBook has like 32gb of RAM 😅

[Frank Bouwens]

Not sure it will work with 8gb

[Stefan Oltmann]

There are good reasons to assume that it won't, because Apple goes to 16 GB because of "Apple Intelligence"

[Stefan Oltmann]

On the other hand rumor has it that deepseek runs on raspberries

[Robert Jaros]

Just start with the smallest version. If it works try to run bigger one.

[Stefan Oltmann]

My PC is i7-9700K with 64 GB RAM & RTX 2080 - I guess I try that first.

[Stefan Oltmann]

Agreed. I don't find any numbers, so I will have to test it on my own to see what works.

[Robert Jaros]

But the difference between 1.5b (the smallest) and 32b is huuuge. The smallest one doesn't even speak my native language (Polish) 😉

[ade]

Rtx 2080 can easily run ollama deepseek-r1:8b

[Stefan Oltmann]

Thanks, that was what I was looking for :)

[Robert Jaros]

And completely off-topic 😉

[Stefan Oltmann]

Nah 😛

[Stefan Oltmann]

I test it and will write something regarding the Compose UI. We'll get back on topic, don't worry 😄

[Stefan Oltmann]

Regarding the installation step ollama run deepseek-r1:1.5b I wonder if that's only available via CLI or if the HTTP-connected
OllamaService could also do that... In that case I'd wish for a GUI to search and select models in the app.
Just so the steps are reduced to "install Ollama & start the app"

[Stefan Oltmann]

There is an unused pullModel() which makes me believe it's possible & planned

[Stefan Oltmann]

@Kashismails Maybe add a "refresh" button to the "Select Model" list 😉

[Stefan Oltmann]

UX/UI Feedback:

Padding Consistency: Reduce the padding around the textfield to a consistent8dp (as shown in the screenshot). Double padding should only be used in rare cases where additional space is necessary for visual hierarchy or content clarity.

Keyboard Functionality: Ensure that pressing the Enter key sends the input directly, while pressing Shift + Enter moves the cursor to the next line. This behavior aligns with common practices seen in applications like ChatGPT and other text inputs.

Copy & Paste of text: Try to copy the style / formattings, like a list.

[Stefan Oltmann]

I hope that's on-topic enough. 😄

8b is really fast on RTX 2080. I try to go higher.

[Stefan Oltmann]

Ah, and offer in the UI an option to still show that <think> part after the answer is given. That's interesting.

[Stefan Oltmann]

I would just try to copy https://chat.deepseek.com as close as possible, I guess.

[Stefan Oltmann]

On a side note:
On RTX 2080 the 32b is the highest that runs, above it just crashes. But it takes forever. 8b was fast. 14b is "acceptable", but I think I stay with 8b.
My MacBook Air M1 8GB can run the 8b version, but answering speed is like the first version of ChatGPT and similar to 14b on RTX 2080.

[Kashismails]

thanks for the feedback @Stefan Oltmann i did face the issue but was lazy to fix it 😛 will fix this today

[Kashismails]

jumping on a discussion i have an m1 max 32gb it runs 14b really good and m3air 16gb which can run 7b but the speed isnt very good