I really like Jan, especially the organization's principles: https://jan.ai/
Main deal breaker for me when I tried it was I couldn't talk to multiple models at once, even if they were remote models on OpenRouter. If I ask a question in one chat, then switch to another chat and ask a question, it will block until the first one is done.
Also Tauri apps feel pretty clunky on Linux for me.
diggan
> Also Tauri apps feel pretty clunky on Linux for me.
All of them, or this one specifically? I've developed a bunch of tiny apps for my own usage (on Linux) with Tauri (maybe largest is just 5-6K LoC) and always felt snappy to me, mostly doing all the data processing with Rust then the UI part with ClojureScript+Reagent.
signbcc
> especially the organization's principles
I met the team late last year. They’re based out of Singapore and Vietnam. They ghosted me after promising to have two follow-up meetings, and were unresponsive to any emails, like they just dropped dead.
Principles and manifestos are a dime a dozen. It matters if you live by them or just have them as PR pieces. These folks are the latter.
dcreater
With a name like Menlo research, I assumed they were based in Menlo park. They probably intended that
_the_inflator
Yep. I really see them as an architecture blueprint with a reference implementation and not so much as a one size fits all app.
I stumbled upon Jan.ai a couple of months ago when I was considering a similar app approach. I was curious because Jan.ai went way beyond what I considered to be limitations.
I haven’t tried Jan.ai yet, I see it as an implementation not a solution.
inkyoto
> Main deal breaker for me when I tried it was I couldn't talk to multiple models at once […]
… which seems particularly strange considering the size of the cloned GitHub repository to be 1.8GiB which swells up to 4.8GiB after running «make build» – I tried to build it locally (which failed anyway).
It is startling that a relatively simple UI frontend can add 3Gb+ of build artefacts alone – that is the scale of a Linux kernel build.
c-hendricks
Yeah, webkit2gtk is a bit of a drag
roscas
Tried to run Jan but it does not start llama server. It also tries to allocate 30gb that is the size of the model but my vram is only 10gb and machine is 32gb, so it does not make sense. Ollama works perfect with 30b models.
Another thing that is not good is that it make constant connections to github and other sites.
hoppp
It probably loads the entire model into ram at once while ollama solves this and does not, it has a better loading strategy
blooalien
Yeah, if I remember correctly, Ollama loads models in "layers" and is capable of putting some layers in GPU RAM and the rest in regular system RAM.
trilogic(dead)
[flagged]
woadwarrior01
> If you looking for privacy there is only 1 app in the whole wide internet right now, HugstonOne
That's a tall claim.
I've been selling a macOS and iOS private LLM app on the App Store for over two years now, that is:
a) is fully native (not electron.js)
b) not a llama.cpp / MLX wrapper
c) fully sandboxed (none of Jan, Ollama, LM Studio are)
I will not promote. Quite shameless of you to shill your electron.js based llama.cpp wrapper here.
threecheese
Purchased, to show my support (and to play around ofc).
woadwarrior01
Thanks! Also check out my other free privacy focussed app. :)
The app has a licence well visible when you install the app. The rest is written in the website and in github. Then about: requires an activation code to use, ofc it is made for ethical research purposes, so yes I am distributing it responsibly. And you can see the videos in the youtube channel for how it works. But the most important point is that you can try it easily with a firewall to see that it do not leak bytes like all the rest there. That´s what i call privacy, It has a button that cut all connection. You can say what you want but that´s it that´s all.
do_not_redeem
> But the most important point is that you can try it easily with a firewall to see that it do not leak bytes like all the rest there.
Great to hear! Since you care so much about privacy, how can I get an activation code without sending any bytes over a network or revealing my email address?
riquito
Closed source, without 3rd party independent review and people should just trust you? As if your app cannot start sending data away in a month or attempt to detect monitoring software, to name a couple
kgeist
>I challenge everyone to find another local GUI with that privacy
Llama.cpp's built-in web UI.
trilogic
This is from webui website docs: Once saved, Open WebUI will begin using your local Llama.cpp server as a backend!
So you see Llama server not CLI. That´s a big flag there. I repeat no app in the whole world takes seriously privacy like HugstonOne. This is not advertisement, I am just making a point.
kgeist
I'm not sure what you're talking about. Llama.cpp is an inference server which runs LLMs locally. It has a built-in web UI. You can't get more private than the inference server itself.
I tried downloading your app, and it's a whopping 500 MB. What takes up the most disk space? The llama-server binary with the built-in web UI is like a couple MBs.
Did you see the feature list? It does not deny that makes connections to other sites.
- Cloud Integration: Connect to OpenAI, Anthropic, Mistral, Groq, and others
- Privacy First: Everything runs locally when you want it to
mathfailure
Is this an alternative to OpenWebUI?
apitman
Not exactly. OWUI is a server with a web app frontend. Jan is a desktop app you install. But it does have the ability to run a server for other apps like OWUI to talk to.
ekianjo
Openweb-ui does not include a server.
cristoperb
It starts a webserver to serve its UI, which is what your comment parent meant. It doesn't provide its own openai-style API, which I guess is what you meant.
apitman
I was referring to Jan.
PeterStuer
More an alternative to LM Studio I think from the description.
apitman
Jan also supports connecting to remote APIs (like OpenRouter), which I don't think LM Studio does
jwildeboer
My name is Jan and I am not an AI thingy. Just FTR. :)
underlines
Jan here too, and I work with LLMs full time and I'm a speaker about these topics. Annoying how many times people ask me if Jan.ai is me lol
dsp_person
We need a steve.ai
ithkuil
I want a Robert Duck AI
tough
We're the AI's Robert's Ducks
ithkuil
nice one.
we're AI's fitness function
klausa
So this is how women named Siri felt in 2011.
lagniappe
Hello Jan ;)
biinjo
Im confused. Isn’t the whole premise of Ollama that its locallt ran? What’s the difference or USP when comparing the two.
moron4hire
That's not the actual tagline being used in the repo. The repo calls itself an alternative to ChatGPT. Whoever submitted the link changed it.
hoppp
I think its an alternative because ollama has no UI and its hard to use for non-developers who will never touch the CLI
I have been using the Ollama GUI on Windows since release and appreciated its simplicity. It recently received an update that puts a large "Turbo" button in the message box that links to a sign-in page.
I'm trying Jan now and am really liking it - it feels friendlier than the Ollama GUI.
dcreater
And ollamas founder was on here posting that they are still focused on local inference... I don't see ollama as anything more than a funnel for their subscription now
numpad0
I truly don't understand why it's supposed to be the end of the world. They need to monetize eventually, and simultaneously its userbase desireg good inference. It looks a complete win-win to me. Anyone can fork it in case they actually turn evil once it'd happen.
I mean, it's not like people enjoy lovely smell of cash burning and bias opinions heavily towards it, or is it like that?
reader9274
Tried to run the gpt-oss:20b in ollama (runs perfectly) and tried to connect ollama to jan but it didn't work.
accrual
I got Jan working with Ollama today. Jan reported it couldn't connect to my Ollama instance on the same host despite it working fine for other apps.
I captured loopback and noticed Ollama returning an HTTP 403 forbidden message to Jan.
The solution was set environment variables:
OLLAMA_HOST=0.0.0.0
OLLAMA_ORIGINS=*
Here's the rest of the steps:
- Jan > Settings > Model Providers
- Add new provider called "Ollama"
- Set API key to "ollama" and point to http://localhost:11434/v1
- Ensure variables above are set
- Click "Refresh" and the models should load
Note: Even though an API key is not required for local Ollama, Jan apparently doesn't consider it a valid endpoint unless a key is provided. I set mine to "ollama" and then it allowed me to start a chat.
still looking for vLLM to support Mac ARM Metal GPUs
baggiponte
Yeah. The docs tell you that you should build it yourself, but…
tough
but unlike cuda there's no custom kernels for inference in vllm repo...
I think
bogdart
I tried Jan last year, but the UI was quite buggy. But maybe they fixed it.
diggan
Please do try it out again, if things used to be broken but they no longer are, it's a good signal that they're gaining stability :) And if it's still broken, even better signal that they're not addressing bugs which would be worse.
esafak
So you're saying bugs are good?!
diggan
No, but maybe that their shared opinion will be a lot more insightful if they provide a comparison between then and now, instead of leaving it at "it was like that before, now I don't know".
venkyvb
How does this compare to LM studio ?
rmonvfer
I use both and Jan is basically the OSS version of LM Studio with some added features (e.g, you can use remote providers)
I first used Jan some time ago and didn’t really like it but it has improved a lot so I encourage everyone to try it, it’s a great project
angelmm
For me, the main difference is that LM Studio main app is not OSS. But they are similar in terms of features, although I didn't use LM Studio that much.
Main deal breaker for me when I tried it was I couldn't talk to multiple models at once, even if they were remote models on OpenRouter. If I ask a question in one chat, then switch to another chat and ask a question, it will block until the first one is done.
Also Tauri apps feel pretty clunky on Linux for me.
All of them, or this one specifically? I've developed a bunch of tiny apps for my own usage (on Linux) with Tauri (maybe largest is just 5-6K LoC) and always felt snappy to me, mostly doing all the data processing with Rust then the UI part with ClojureScript+Reagent.
I met the team late last year. They’re based out of Singapore and Vietnam. They ghosted me after promising to have two follow-up meetings, and were unresponsive to any emails, like they just dropped dead.
Principles and manifestos are a dime a dozen. It matters if you live by them or just have them as PR pieces. These folks are the latter.
I stumbled upon Jan.ai a couple of months ago when I was considering a similar app approach. I was curious because Jan.ai went way beyond what I considered to be limitations.
I haven’t tried Jan.ai yet, I see it as an implementation not a solution.
… which seems particularly strange considering the size of the cloned GitHub repository to be 1.8GiB which swells up to 4.8GiB after running «make build» – I tried to build it locally (which failed anyway).
It is startling that a relatively simple UI frontend can add 3Gb+ of build artefacts alone – that is the scale of a Linux kernel build.
That's a tall claim.
I've been selling a macOS and iOS private LLM app on the App Store for over two years now, that is:
a) is fully native (not electron.js) b) not a llama.cpp / MLX wrapper c) fully sandboxed (none of Jan, Ollama, LM Studio are)
I will not promote. Quite shameless of you to shill your electron.js based llama.cpp wrapper here.
> I accept every challenge to prove that HugstonOne is worth the claim.
I expect your review.
It's not open source, has no license, runs on Windows only, and requires an activation code to use.
Also, the privacy policy on their website is missing[2].
Anyone remotely concerned about privacy wouldn't come near this thing.
Ah, you're the author, no wonder you're shilling for it.
[1]: https://github.com/Mainframework/HugstonOne
[2]: https://hugston.com/privacy
Great to hear! Since you care so much about privacy, how can I get an activation code without sending any bytes over a network or revealing my email address?
Llama.cpp's built-in web UI.
I tried downloading your app, and it's a whopping 500 MB. What takes up the most disk space? The llama-server binary with the built-in web UI is like a couple MBs.
- Cloud Integration: Connect to OpenAI, Anthropic, Mistral, Groq, and others
- Privacy First: Everything runs locally when you want it to
we're AI's fitness function
I'm trying Jan now and am really liking it - it feels friendlier than the Ollama GUI.
I mean, it's not like people enjoy lovely smell of cash burning and bias opinions heavily towards it, or is it like that?
I captured loopback and noticed Ollama returning an HTTP 403 forbidden message to Jan.
The solution was set environment variables:
Here's the rest of the steps:- Jan > Settings > Model Providers
- Add new provider called "Ollama"
- Set API key to "ollama" and point to http://localhost:11434/v1
- Ensure variables above are set
- Click "Refresh" and the models should load
Note: Even though an API key is not required for local Ollama, Jan apparently doesn't consider it a valid endpoint unless a key is provided. I set mine to "ollama" and then it allowed me to start a chat.
Can't make it work with ollama endpoint
this seems to be the problem but they're not focusing on it: https://github.com/menloresearch/jan/issues/5474#issuecommen...
I think
I first used Jan some time ago and didn’t really like it but it has improved a lot so I encourage everyone to try it, it’s a great project