LM Link

Posted by Blindax@reddit | LocalLLaMA | View on Reddit | 40 comments

I see that LM Studio just shadow dropped one of the most amazing features ever. I have been waiting this for a long time. LM Link allows a client machine to connect to another machine acting as server remotely using tailscale. This is now integrated in the LM Studio app (which either acts as server or client) and using the GUI. Basically, this means you can now use on your laptop all your models present on your main workstation/server just as if you were sitting in front of it. The feature is currently included in the 0.4.5 build 2 that just released and it's in preview (access needs to be requested and is granted in batches / i got mine minutes after request). It seems to work incredibily well. Once again these guys nailed it. Congrats to the team!!!

Reply to Post

40 Comments

[-]

sturmen@reddit

My dream is that they’re also cooking up native smartphone apps so I can use my local LLMs on my phone just the same as the ChatGPT or Claude apps

[-]

TacGibs@reddit

Just use Conduit (it's open source and on the playstore), llama.cpp (or ikllamacpp or vLLM or whatever) and OpenwebUI.

[-]

sturmen@reddit

Different strokes for different folks, but I like using LM Studio and I’m hopeful that a smartphone app is on their roadmap.

[-]

VideoGameJumanji@reddit

LM mini is a free app on iOS that is pretty much 1:1 with LM studio, try it if you haven't the only thing I'd really like to see rolled out is chat history being shared over cloud somehow

[-]

iMrParker@reddit

The fact that it requires an account is a no-go for me

[-]

Sazaju@reddit

Also a no-go for me. I want them to connect without intermediaries: just provide the IP/port of one instance in the other instance, request the connection, accept it on the other side, and that should suffice to communicate in both ways.

[-]

soostenuto@reddit

I don't get it. Why do I need a fucking account to connect to my own remote server wtf??

[-]

derivative49@reddit

instead they could've just used a combination for the account with a password for client confirmation and this would've been sorted

[-]

derivative49@reddit

i know right? sounds like a way to start profiling your activity

[-]

_-_David@reddit

Can someone ehlp me understand the excitement? I use my LM Studio server constantly, but this would essentially mean I can use the LM Studio app from another pc, in say, another state? That doesn't sound particularly great. I've never had one interaction of meaning with the chat section of the app. Is this a a secure way of extending the local server to clients outside the local network? I want to get excited, I just don't grok the significance of this.

[-]

Blindax@reddit (OP)

If you use the models exclusively for chat, the server is fine. If you want to be able to tweak parameters remotely this is where it becomes painful. I know I change things like context windows constantly depending on the size of the documents I want to analyse. This can be done easily locally or remotely with lm link. As far as I know there was not an easy way to do it before with the server, at least not with a nice gui.

[-]

_-_David@reddit

And do things like toggle reasoning?! If you're saying I can use this the same way I've been using the local server, but with increased flexibility I'll be chuffed. For instance, it's been driving me a little crazy lately that I can't seem to get LM Studio to allow a model to reason without structured outputs, then apply the json schema during the non-reasoning section. I'm having to use hacky tricks like parsing the structured output straight out of the reasoning trace, and adding a prefix to the schema that requires reasoning as an output in that json. Or making a call and asking for reasoning, then a second call and asking for structured output from the total context. Any sort of configurability features that I can toggle remotely are well appreciated. Thanks for the info

[-]

Blindax@reddit (OP)

I have not spent too much time on it. But the issue with the server (at least from my understanding) is that your model with be served « as is » and setting in the interface you use (say open web ui) won’t work. Here you use the lm studio interface like with a local model and all settings are available (+ the api point is also served locally) including reasoning toggles I expect for models that support it.

[-]

_-_David@reddit

Okay, so it really is basically "Open your local LM Studio from anywhere" more than anything.

[-]

CallumCarmicheal@reddit

Its more of remote processing, your chats are not synced / shared as they would be with a web ui server like OpenWebUI. You will still need a external system to sync your chats if you wish for that. This is nothing more then a way to link your LM Studio frontend/gui to a remote backend where all the AI processing is done remotely then sent back. Its like a openai api backend but with tighter integration to the remote UI.

[-]

derivative49@reddit

The "necessity" of signing in prevents me from using this. what's the point when it could be done without accounts? why hide features behind sign ups? i thought this was the one thing we weren't meant to be doing...

[-]

maxwell_aws@reddit

How do I call remote host in API? Both local and remote have the same model, I'm looking to parallelize execution

[-]

Common_Heron4002@reddit

I'm over the moon about this on main source applications. Anyone know why we are required to sign in....why there is no just local use for it without sign in ?

[-]

Hot-Nefariousness497@reddit

It was looking great this morning. Then it stopped connecting. Am I the only one?

[-]

No_Conversation9561@reddit

Hope IOS and Android app also are in their plan.

[-]

floppypancakes4u@reddit

How is this different than tailgate? Just more convenient?

[-]

Blindax@reddit (OP)

Not sure what tailgate is but if you mean tailscale, this means you don’t need to log in a remote os session so it is indeed much easier to use.

[-]

floppypancakes4u@reddit

Yes I meant tailscale. What do you mean log in to a remote session? Are you using the lmstudio chat interface itself to use models?

[-]

Blindax@reddit (OP)

Yes. In your model list you will have your local models + those accessible through lm link in remote machine

[-]

FPham@reddit

Wait, why do I need to use tailscale if I'm on my own network? Why does it needs to go through them? Or am I not getting it?

[-]

AnticitizenPrime@reddit

About time. I actually ditched LM Studio for Msty + tailscale a year and a half or so go because I was annoyed that I couldn't use LM Studio as a remote client for my desktop server. Msty has done both from the start (though you have to set up Tailscale on your own, but it's easy).

[-]

rm-rf-rm@reddit

What did Msty do exactly? Any inference engine that exposes a REST API is all you need to then access it remotely through Tailscale

[-]

AnticitizenPrime@reddit

It's basically the same thing as LM Studio in that it's a GUI that makes using local models easily. It's one GUI app that is easy to install and has its inference engine built in and creates a server, like LM Studio. It originally had just Ollama built in but lately has an option for Llama.cpp as well. It's attractive for the same reasons that LM Studio is, no futzing around with command line stuff, just install an app and away we go. The difference is that (until now) LM Studio was only for use on one machine. Sure it provided a server, but to access it remotely you'd need a completely different app if you were on your laptop, say, and wanted to access your models that were on your desktop. You'd have to use OpenWebUi or even Msty or whatever. You couldn't use LM Studio to connect to your own LM Studio instance. Msty does what LM Studio has done in the past (the actual management of hosting models locally), but also has always done what LM Studio is just now doing, which is to serve as a remote client, so you only need to learn one app. You can also use Msty with any API out there too, as well as with your local LLMs, so it's an all-in-one client, as well as server. As a client, it's also pretty damn feature rich, so much so that I've never learned to use them all. It had RAG and MCP stuff baked in before LM Studio did. I'm actually behind on all its features, and haven't migrated fully to the new version (Msty Studio) which has a ton more.

[-]

rm-rf-rm@reddit

I know what Msty is... I dont think its right to market it as "can be accessed remotely" as that functionality is just literally from llama.cpp (which ollama wrapped and then msty wraps on top of that). For casual users, I advise against using Msty. Though its what I use (because I cant find anything that is sufficiently better to move), its underbaked in its engineering and UI polish (i have no idea what stack they are using but its the strangest one i've ever seen and gives me an ick) and overbaked in toy implementations of a bunch of things - MCP, RAG, "Agents" etc. To the point I don't use any of those at all as I have no faith that it works well. The whole pivot to Msty Studio is a clear tell of their enshittification journey starting.

[-]

AnticitizenPrime@reddit

The pivot to Studio is because they decided to rebuild from the ground up rather than re-patch their original application. I have no idea why you think that's enshittification. >I advise against using Msty. Though its what I use (because I cant find anything that is sufficiently better to move) Well, that is certainly a take.

[-]

Blindax@reddit (OP)

I never tested it, I was either using open web ui or tailscaling to a rdp instance with lm studio which was not the most convenient. What is very nice is that with LM link is that you get you can keep all the granularity of model settings with the client. I always find it very painful to manage items like context window, gpu offload or this kind of things with Open Web UI.

[-]

PieBru@reddit

What's the difference vs directly using 'llama-server --host 0.0.0.0' via Tailscale?

[-]

mantafloppy@reddit

Its in the release note : > 0.4.5 - Release Notes > Build 2 > > Fixed a bug where LM Link connector was not included in in-app updater > Build 1 > > ✨🎉 Introducing LM Link > Connect to remote instances of LM Studio, load your models, and use them as if they were local. > End-to-end encrypted. Launching in partnership with Tailscale. > Improved tool calling support for the Qwen 3.5 model family > Fixed a bug where loading model would sometimes fail with "Attempt to pull a snapshot of system resources failed. Error: 'Utility process is not defined'". > Fixed a bug where autoscrolling new message behavior was not respected when clicking the Generate button > Hides the Generate button when editing a message to avoid accidental click

[-]

Blindax@reddit (OP)

I was doing exactly the same, this will be much more convenient :)

[-]

HopePupal@reddit

oh finally. LM Studio's UI is much more reliable than AnythingLLM's. i'd started looking into web UIs but this sounds a little more convenient