7b is the smallest I've found useful. I'd try a smaller quant before going lower, if I had super small vram.
this post was submitted on 16 Jun 2025
85 points (94.7% liked)
Selfhosted
46672 readers
1385 users here now
A place to share alternatives to popular online services that can be self-hosted without giving up privacy or locking you into a service you don't control.
Rules:
-
Be civil: we're here to support and learn from one another. Insults won't be tolerated. Flame wars are frowned upon.
-
No spam posting.
-
Posts have to be centered around self-hosting. There are other communities for discussing hardware or home computing. If it's not obvious why your post topic revolves around selfhosting, please include details to make it clear.
-
Don't duplicate the full text of your blog or github here. Just post the link for folks to click.
-
Submission headline should match the article title (don’t cherry-pick information from the title to fit your agenda).
-
No trolling.
Resources:
- selfh.st Newsletter and index of selfhosted software and apps
- awesome-selfhosted software
- awesome-sysadmin resources
- Self-Hosted Podcast from Jupiter Broadcasting
Any issues on the community? Report it using the report flag.
Questions? DM the mods!
founded 2 years ago
MODERATORS
It'll work for quick bash scripts and one-off things like that. But there's not usually enough context window unless you're using a 24G GPU or such.
Snippets are a great use.
I use StableCode on my phone as a programming tutor for learning Python. It is outstanding in both speed and in accuracy for this task. I have it generate definitions which I copy and paste into Anki the flashcard app. Whenever I'm on a bus or airplane I just start studying. Wish that it could also quiz me interactively.
I installed Llama. I've not found any use for it. I mean, I've asked it for a recipe because recipe websites suck, but that's about it.
you can do a lot with it.
I heated my office with it this past winter.
I've run a few models that I could on my GPU. I don't think the smaller models are really good enough. They can do stuff, sure, but to get anything out of it, I think you need the larger models.
They can be used for basic things, though. There are coder specific models you can look at. Deepseek and qwen coder are some popular ones
Been coming to similar conclusions with some local adventures. It's decent but not as able to process larger contexts.
Have you tried RAG? I believe that they are actually pretty good for searching and compiling content from RAG.
So in theory you could have it connect to all of you local documents and use it for quick questions. Or maybe connected to your signal/whatsapp/sms chat history to ask questions about past conversations
No, what is it? How do I try it?
RAG is basically like telling an LLM "look here for more info before you answer" so it can check out local documents to give an answer that is more relevant to you.
You just search "open web ui rag" and find plenty kf explanations and tutorials
I think RAG will be surpassed by LLMs in a loop with tool calling (aka agents), with search being one of the tools.
LLMs that train LoRas on the fly then query themselves with the LoRa applied
Sorry, I am just gonne dump you some links from my bookmarks that were related and interesting to read, cause I am traveling and have to get up in a minute, but I've been interested in this topic for a while. All of the links discuss at least some usecases. For some reason microsoft is really into tiny models and made big breakthroughs there.
https://reddit.com/r/LocalLLaMA/comments/1cdrw7p/what_are_the_potential_uses_of_small_less_than_3b/
https://github.com/microsoft/BitNet
https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/
https://news.microsoft.com/source/features/ai/the-phi-3-small-language-models-with-big-potential/
I've used smollm2:135m for projects in DBeaver building larger queries. The box it runs on is Intel HD graphics with an old Ryzen processor. Doesn't seem to really stress the CPU.
UPDATE: I apologize to the downvoter for not masochistically wanting to build a 1000 line bulk insert statement by hand.
Converting free text to standardized forms such as json
Oh—do you happen to have any recommendations for that?
DeepSeek-R1-Distill-Qwen-1.5B
I've integrated mine into Home Assistant, which makes it easier to use their voice commands.
I haven't done a ton with it yet besides set it up, though, since I'm still getting proxmox configured on my gaming rig.
What are you using for voice integration? I really don't want to buy and assemble their solution if I don't have to
I just use the companion app for now. But I am designing a HAL9000 system for my home.
Currently I've been using a local AI (a couple different kinds) to first - take the audio from a Twitch stream; so that I have context about the conversation, convert it to text, and then use a second AI; an LLM fed the first AIs translation + twitch chat and store 'facts' about specific users so that they can be referenced quickly for a streamer who has ADHD in order to be more personable.
That way, the guy can ask User X how their mothers surgery went. Or he can remember that User K has a birthday coming up. Or remember that User G's son just got a PS5 for Christmas, and wants a specific game.
It allows him to be more personable because he has issues remembering details about his users. It's still kind of a big alpha test at the moment, because we don't know the best way to display the 'data', but it functions as an aid.
Hey, you're treating that data with the respect it demands, right? And you definitely collected consent from those chat participants before you Hoover'd up their [re-reads example] extremely Personal Identification Information AND Personal Health Information, right? Because if you didn't, you're in violation of a bunch of laws and the Twitch TOS.
If I say my name is Doo doo head, in a public park, and someone happens to overhear it - they can do with that information whatever they want. Same thing. If you wanna spew your personal life on Twitch, there are bots that listen to all of the channels everywhere on twitch. They aren't violating any laws, or Twitch TOS. So, *buzzer* WRONG.
Right now, the same thing is being done to you on Lemmy. And Reddit. And Facebook. And everywhere else.
Look at a bot called "FrostyTools" for Twitch. Reads Twitch chat, Uses an AI to provide summaries of chat every 30 minutes or so. If that's not violating TOS, then neither am I. And thousands upon thousands of people use FrostyTools.
I have the consent of the streamer, I have the consent of Twitch (through their developer API), and upon using Twitch, you give the right to them to collect, distribute, and use that data at their whim.
So, buzzer WRONG.
Quite arrogant after you just constructed a faulty comparison.
If I say my name is Doo doo head, in a public park, and someone happens to overhear it - they can do with that information whatever they want. Same thing.
That's absolutely not the same thing. Overhearing something that is in the background is fundamentally different from actively recording everything going on in a public space. You film yourself or some performance in a park and someone happens to be in the background? No problem. You build a system to identify everyone in the park and collect recordings of their conversations? Absolutely a problem, depending on the jurisdiction. The intent of the recording(s) and the reasonable expectations of the people recorded are factored in in many jurisdictions, and being in public doesn't automatically entail consent to being recorded.
See for example https://www.freedomforum.org/recording-in-public/
(And just to clarify: I am not arguing against your explanation of Twitch's TOS, only against the bad comparison you brought.)
You build a system to identify everyone in the park and collect recordings of their conversations? Absolutely a problem, depending on the jurisdiction.
Literally not. The police use this right now to record your location and time seen using license plates all over the nation - with private corporations providing the service.
and being in public doesn't automatically entail consent to being recorded.
And yes, it's called 'expectation to the right of privacy'. Public venues are not 'private' locations, and thus do not need consent. You can, quite literally, record anyone in public.
Even the link you provided agrees.
Doesn't Twitch own all data that is written and their TOS will state something like you can't store data yourself locally.
I'm not storing their data. I'm feeding it to an LLM which infers things and storing that data.
Most US states are single party consent. https://recordinglaw.com/united-states-recording-laws/one-party-consent-states/
There is no expectation of privacy in public spaces. Participants to these streams which are open to all do not have a prohibition on repeating what they have heard.
Right and what I was saying was even if it wasnt “public”, single party consent means the person recording can be that single party- so still a non-issue.
Surely none of that uses a small LLM <= 3B?
Yes. The small LLM isn't retrieving data, it's just understanding context of text enough to know what "Facts" need to be written to a file. I'm using the publicly released Deepseek models from a couple of months ago.
sounds like salesforce for a twitch setting. cool use case, must make fun moments when he mentions such things.
Esp. if the LLM just hallucinates 50% of the "facts" a about the users 👌
That hasn't been a problem at all for the 200+ users it's tracking so far for about 4 months.
I don't know a human that could ever keep up with this kind of thing. People just think he's super personable, but in reality he's not. He's just got a really cool tool to use.
He's managed some really good numbers because being that personal with people brings them back and keeps them chatting. He'll be pushing for partner after streaming for only a year and he's just some guy I found playing Wild Hearts with 0 viewers one day... :P
I think that's a size where it's a bit more than a good autocomplete. Could be part of a chain for retrieval augmented generation. Maybe some specific tasks. And there are small machine learning models that can do translation or sentiment analysis, though I don't think those are your regular LLM chatbots... And well, you can ask basic questions and write dialogue. Something like "What is an Alpaca?" will work. But they don't have much knowledge under 8B parameters and they regularly struggle to apply their knowledge to a given task at smaller sizes. At least that's my experience. They've become way better at smaller sizes during the last year or so. But they're very limited.
I'm not sure what you intend to do. If you have some specific thing you'd like an LLM to do, you need to pick the correct one. If you don't have any use-case... just run an arbitrary one and tinker around?
absolutely nothing
Nothing.