Run ANY Open-Source LLM Locally (No-Code LMStudio Tutorial)

LMStudio tutorial and walkthrough of their new features: multi-model support (parallel and serialized) and JSON outputs.
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewberman.com
Need AI Consulting? ✅
forwardfuture.ai/
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Rent a GPU (MassedCompute) 🚀
bit.ly/matthew-berman-youtube
USE CODE "MatthewBerman" for 50% discount
Media/Sponsorship Inquiries 📈
bit.ly/44TC45V
Links:
LMStudio - lmstudio.ai/
LMStudio Tutorial 1 - • Run ANY Open-Source Mo...
Disclosures:
I'm an investor in LMStudio and CrewAI

[@bonkywonky1]

Project idea: create a bunch of agents that are experts in specific areas, like coding, Wikipedia, reasoning, law, etc, and then an orchestrating agent. The orchestrator will be the only one the user then interacts with. The orchestrator then figures out how to respond to user queries by finding all available agents available and selecting either one or multiple to produce the best answer possible. Either the agents have descriptions of what it’s good at or if the orchestrator can see each agents metadata and even just recognize what they’re good at by just seeing how they’re setup, that would be even better.

[@positivevibe142]

There are coding copilots/agents out there already, like Pythagora GPT Pilot, Devin, Devika, Auto-GPT, Github copilot, Warp...etc... Personally, I rely heavily on the current Claude 3 Opus, OpenAI ChatGPT 4 looks like a joke next to it! 😅

[@Thedeepseanomad]

Yes. a specialized and optimized LLM will outperform a general model. If one could successfully train custom specializations, having a constellations of models for specialized tasks could result in very high capacity (like a model that is 7B for Cyberpunk story structure, 7B for dialogue in a Cyberpunk setting, 7B for pacing in adventure stories, etc, you could have a a setup running like 28B for only making a cyberpunk stories, but only running a single model at a time.

[@executivelifehacks6747]

I have to say I use Claude 3 Opus as a first choice for AI.

[@bigglyguy8429]

@@positivevibe142 I have both and I probably won't renew Opus as it's not giving me anything GPT doesn't, but GPT can do more

[@agentxyz]

great idea! you could call it "mixture of experts"

[@starblaiz1986]

I've been using LMStudio since your last tutorial on it, and I can attest that it's FANTASTIC to use and takes all of the headache out of setting up local AI's. TIP: In LMStudio's settings you can specify the folder to download AI models to. It's worth getting a small dedicated flash drive to store them on. That way you can play about with them without having to worry about hard drive space as the smallest models are about 5GB and the largest can get into tripple digits. Yes loading them up will take slightly longer, but inferrance won't be affected as that's done entirely from RAM (and if it's too big for the RAM then it will use your main hard drive for VRAM just like it would normally, so having it on external USB flash doesn't affect it).

[@jayr7741]

Hey should I consider perplexity ai pro subscription to analysis my previous year questions of exam or free LMStudio will also be good please reply

[@serikazero128]

does LM studio allow chat with documents? I'm trying to set this up properly, but even Private GPT doesn't work as it used to

[@myandrobox3427]

​@@serikazero128Anything LLM is good for interacting with docs

[@serikazero128]

@@myandrobox3427 thanks, I'll look into it

[@bigglyguy8429]

@@jayr7741 Use Perplexity for now, unless you have a really high-end machine that can run big models (2x 3090 at 24GB each).

[@karankatke]

We need more usecase and practical guides with LM studio. Love your videos. ❤

[@abdelhakkhalil7684]

Thank you for the video, and thank you for disclosing that you are an investor in both LMStudio and CrewAI. I wish you could mention it in the video for better transparency.

[@xtramoist9999]

Agreed. Would be highly regarded.

[@matthew_berman]

I considered this, maybe I should have. I didn’t want to be…show-off-y.

[@qwazy0158]

Are either of these public? Or are both private companies ?

[@TheExcellentVideoChannel]

Thanks Mat, what would we do without you to guide us on this journey!! I started on ollama and ran into some issues that I don't want to solve just yet but it looks like LMStudio is what I need to move to to get around the issues. Nice and timely tutorial/overview.

[@JohnLewis-old]

Always love your reviews. Thanks!

[@OproDarius]

Such an awesome software, can't wait to see the local open source software delivering agent and llms in 5 years, will be such a ride!

[@issiewizzie]

I'm hoping in the future, there will be a way to train an LLM or specialised model more easily for a beginner. Almost iPhone-friendly.

[@Yakibackk]

No hope bro

[@alx8439]

Oobabooga has it in UI already. You just need to have hardware which can cope it

[@lancemarchetti8673]

The system requirements are very helpful. Thanks

[@jackflash6377]

YES to the agents locally. Any interest in checking out Devika? Claim to be open source Devin.

[@matthew_berman]

I spent a good bit of time trying to get it to work but couldn’t. I’ll certainly do a review when I get it working though. Based on its popularity, I suspect it’ll evolve quickly.

[@rudolfviljoen2847]

​@@matthew_berman If you do make a video on devika please please do a section on using it with a local LLM

[@krisknap]

Thanks for sharing this update and demonstrating with the examples!

[@BudoReflex]

I love how you move through topics and keep a concise summary of what is happening without going down rabbit holes. I learn a lot very quickly.

[@matthew_berman]

Thank you!

[@riyadwahib4755]

Thanks Matt! Great video as usual :) Yes please would be nice to see you build something with powering agents!

[@OzzyMoto2K10]

Great video, Matt - you have no longer jumped the shark. :)

[@morena-jackson]

Love this type of video, thanks so much for really going into LMStudio. I"ve had the program for a dew months now but never really played with.

[@bigglyguy8429]

If you're using it for role-play also looking into Faraday, which is more setup for role-play and I find it runs the same models faster.

[@morena-jackson]

​@@bigglyguy8429 thank you!!

[@eagleterry3349]

Looking forward to seeing you build a project .

[@DanielArnolf]

This is getting cooler by the day, what a gift !!! Thanks for your professionalism and dedication.

[@wakingdreamsroleplay]

I am not a coder so I love your videos. This agents stuff excites and terrifies me. I would love to see a project that involves writing complicated documents such as text-based games (for improvised drama type activities) where the character descriptions have clue info about other players. So, the writing agent creates setting and background and character descriptions but an editor needs to go through the full document and check and make sure clues found in one character file also show up in other files and send back to writer to reiterate until everything checks out. I can do part of this with prompting but it always needs a lot of manual editing and so a way to automate would be nice. If you can demonstrate some similar process, I would certainly be thrilled. ;-)

[@vibeymonk]

Loved this video, please make a playlist out of it! More inclined for people like me.

[@panagiotisgalinos1335]

Man, i like your videos. Very informative.

[@armans4494]

Yes, please. Also publish your endpoints for consumers 🎉❤

[@matthew_berman]

Endpoints?

[@supahfly_uk]

Wow this is amazing, thank you.

[@screamingiraffe]

best video so far, thank you for this

[@vladyslavkorenyak872]

Wow, this is Gold! I would love a tutorial on how to integrate this into a website.

[@anubisai]

Great video, Matt.

[@REDULE26]

Nice video as always 👍

[@lucademarco5969]

I would really love to see document q&a using lmstudio, because I think a lot of companies are interested in this kind of ai use.

[@sperazza]

fantastic, really great video

[@neugen1019]

Matt you tricked us yesterday with that 01 thing and that voice distraction

[@erikjohnson9112]

I think that was actually real. The presenter just wants attention, they want to get noticed when they speak (a form of narcissism). I found the product to be interesting and the presentation to be distracting (a negative because it draws away from what is being presented).

[@Akcvs]

@@erikjohnson9112 actually I think you are the narcissist and are just projecting. Transgender people have existed in every region of the Earth since before civilization itself. They're a real naturally occurring demographic. Get over it

[@teddygbg]

Thanks for this Matthew! Super helpful overview.

[@colmxbyrne]

LM Studio really useful

[@johnp9091]

Such a good way to test drive some of the local models. Great job on this and all of your other tutorials! I've really learned a lot from you vids.

[@333dsteele1]

Great video

[@trazercreations8478]

also you can use this with AnythingLLM for documents

[@racerx1777]

DO NOT USE ANYTHINGLLM it is no longer free! Im starting to see a pattern with this video creator. As soon as he releases a video these so called free things all of the sudden become paid for versions! I used AnythingLLM one night on this computer after watching this guys video on AnythingLLM the very next morning at work i went to put it on a computer at work and it was no longer free but subscription based! I AM DONE WITH THE MONEY GAMES BASED ON PRINCIPAL! THEY ARE ALL TRYING TO CASH IN ON THIS CRAP THAT AMOUNTS TO NOTHING MORE THAN A GIMMICK A AI TREND IF YOU WILL!!! BANKRUPT THESE PEOPLE!

[@TomHimanen]

Please demo LM Studio and Crew AI as combo! Also this was a great demo, thanks!

[@agentxyz]

ty. great video

[@dieselphiend]

Considering what I was going through to install models previously, this is dumbfoundingly simple.

[@YusriSalleh]

Excellent video. Tqvm! Is there any video considering various option of hardwares for running local? Say various Nvidia GPUS, or AMD Rocm or even apple metal ?

[@rickevans7941]

How can good content from a genuine, ethical and knowledgeable person only have 200k subs in this niche when DS has 143k with his empty parroting grift? You are headed to ,,, broski, hope you're ready. You deserve it.

[@mikhailkalashnik0v]

Great video thanks! Any known good models I can use for infosec & or application security (pen testing)?

[@aott6799]

Really excellent videos on LMStudio. Does it have the capability to access local files to update chats with data newer than the cut-off date of the LLM? I'd like to be able to input locally stored ebooks and generate summaries along the lines of "What is 'This Book' about?"

[@punishedproduct]

Agents working locally!!❤

[@justinrose8661]

Thanks Matt! You're my favorite AI KZbinr. I'd love to see you build something cool, we all would I'm sure.

[@imacuser101]

use autogen to create a series of agents with their agent builder to run a task

[@tech-vp5xe]

Hey Matt, been following you for a long time awesome work. Does LM Studio allow different models to exist on separate GPU's

[@matthew_berman]

Like if you have multiple GPUs? I don’t think so

[@fruitpnchsmuraiG]

Hey, can you share some advice for an undergrad to get into generative models, where to begin or rather what should i learn to understanding the working of LLMs and playing around with them?

[@drlordbasil]

LM studios and their model server is soooo easy.

[@drlordbasil]

im hoping they add combining and fine tuning functions or at least image/other gen models in future.

[@mathieuboisvert6865]

Fantastic, I was just playing with this today. How would you integrate AnythingLLM with these different models running in parallel and interacting through crewAI?

[@thewatersavior]

branching chat would be a dope chrome plugin

[@carthagely122]

Thanks alot

[@Al-Storm]

I run anythingllm on top of this. It has a nice rag setup.

[@hiddenkirby]

Thanks for this video. I love this development setup. How do you properly serve it all from a cloud?

[@jeffg4686]

They should get a visual node setup like comfy UI but for creating purely llm based apps. Some generate button or whatever spits out the python code (or just runs it) Not that it's hard to write, but some might prefer the node setup.

[@NahFam13]

I'd love to see videos on running specific models. I've been able to run almost every model I've downloaded but I can't seem to get StarCoder or StarCoder2 regardless of the preset I use and I've love to get it running without hallucinations, or looping the same sentence. Another thing I'd love to know is what happened to TheBloke!! I heard he stopped converting models and I'm sure the reason behind has to be epic.

[@sizwemsomi239]

this is fire

[@myandrobox3427]

This is awesome! Thanks for sharing again!! I have quick question... I want to run this on server like you showed, create a nocode app using APIs, and have users access this application. Kinda creating for a local group of users. How do you think this is going to work in terms of machine requirements. Please guide if it's good approach! 🙏🏻

[@alanmckeon8321]

Could you use these to help build an application?

[@aimademerich]

Phenomenal

[@WaqarMurtaza-yk2fh]

I am dropping this comment in case if someone is using this on Linux and have a old machine with built in GPU, I was facing this error where it says: "unable to load any model". A quick solution to that would be unchecking GPU acceleration. Hope this helps!

[@user-ef4df8xp8p]

LMStudio is cool.....Please, make more videos on this tool....

[@Myplaylist892]

Is it possible to set folders within the LLMStudio in order to do some document querying?

[@profitunist5876]

Hi, I was wondering how you'd download gated models, like meta-llama/Meta-Llama-3-8B? There are quantized versions from other authors but I'd prefer to download the actual one released by Meta. Thanks

[@mendthedivide]

what kinda chip does your laptop have Matthew Berman? m1 m2 m3?

[@nannan3347]

If LMStudio had a built in RAG it would be perfection.

[@Mangini037]

Yes pleeeease do a video using AutoAgent tutorial. Thanks.

[@babbagebrassworks4278]

Doesn't work on my Pi5 but Ollama does. They only seem to use memory when answering a prompt, so I can multiple version of Ollama running with different models, as long as I only prompt one at a time. AMD Ryzen AI and Intel Core Ultra have NPUs onboard now so no need for big GPU card.

[@faultyogi]

👌"I have learned a lot."

[@rodrimora]

Does it support exl2 quants? If not that would be the only thing imssing to make the switch from textgen web ui

[@b.j.freedman]

I've used LM Studio (and also Jan) with some fairly large models, but...I haven't seen an explanation on how to set up multiple languages (and alphabets if necessary). That should be easy, geeks please let me know. The second question is: where I can find an app / code whatever so the interface is not text-based, but includes speech, both listening and speaking (again, in a few languages), or what I really want-- a streaming avatar (like at HeyGen or D-ID or others) which will listen and then more or less instantly move lips and face, syncing with a spoken answer in the same language as the question. I ask because I want to do this all locally--it's for education and we want to keep it cheap and off line, so middle school students can't go "elsewhere". Suggestions?

[@b.j.freedman]

I have to add that I've created "Assistants" at Hugging Face that can do multiple languages, but it is still text-based. And on line.

[@Paulina-ds5bg]

Can someone suggest the way/tool when can I train open source model with custom data? For example with pdf.? And giving a questions in terms the given data?

[@dreamyrhodes]

Wow finally an UI with documentation? I always hated that they throw all that GGUF, GPTQ, 4Q 5Q _K_M_S at you without ever telling what it means and what it needs to run.

[@michai333]

I can’t believe how fast it’s running, fully offloaded, on my GPU (OC 4090). I’m using Mistral Claud merged 7B Q8.

[@KaiPhox]

There are many models of Grok to download. Do I need all the files or is there a one click download for LM studio?

[@therighteousagent]

will there be a memory and roleplay function implemented?

[@Parisneo]

LM studio is a nice project. The only complaint I have is that it is not open source. People can use LM Studio along with lollms as it can be run as a server and they seem to get very good output. So yeah, this is a very cool and useful tool.

[@mydogsbutler]

One thing which would be useful... how to sync models between LM studio and Ollama (to avoid duplication to save space). LM Studios defaults Local models in windows is: C:\Users\User\.cache\lm-studio\models When I point it to Ollama's in Windows 11 WSL2 ubuntu 22.04 it doesn't work ( \\wsl.localhost\Ubuntu-22.04\|usr\share\ollama\.ollama\models) Anyone know the answer? Is it even possible?

[@Brax1982]

I don't think that's how the compatibility filter works. I checked for a couple models that are in a repo with tensor files. They did not show up if that filter was active. I guess they include "does not work on LM Studio" as being not compatible with your system. Because it isn't compatible with their tool? At first I thought that it's not true that every model on HF could be found on there. But for those that were missing, I then saw that they show up if that filter is off. Of course, this is a sample size of trying 3 models or something. May be outliers. But do any models that are not GGUF work on LM Studio?

[@chrisb9045]

Hi, it is possible to upload to LM studio your own document files ex excel files, photos, pdf, txt?

[@irom77]

Hey, what laptop would you recommend for this stuff ?

[@juanjesusligero391]

While you can run open source LLMs with LMStudio, LMStudio itself is *NOT* open source. I think you should probably clarify that in the video title, since it's a bit misleading.

[@Cross-CutFilms]

I've re-read the title and caption multiple times since reading this comment and I'm confused as to what you find misleading? *Btw this is not an attack.*

[@urphakeandgey6308]

Unless he didn't mention it at all in the video, I think this is being ridiculously pedantic.

[@carstenli]

I switched to Ollama for that reason. It's open source and works like a charm.

[@juanjesusligero391]

​@@Cross-CutFilms No problem, I didn't feel attacked at all ^^ Which I mean is that it can confuse people that read the title too fast and fail to understand that the Open-Source part refers to the models and not to LLMStudio (Like myself XD). When I read the title I thought: "Wait, so they open-sourced LMStudio? That's great! :D"... And then I visited their page looking for a repo, came back here and I realized I didn't read it right ^_^U I think it's very easy to understand the title wrongly as I did (even more for non native English speakers), hence my comment.

[@juanjesusligero391]

@@urphakeandgey6308If you meant my comment, sorry, I didn't want to be pedantic. I just wanted to clarify what the title wanted to say it's open source, since I got it wrong and there could be more people who also misunderstood the title.

[@thegooddoctor6719]

So the big question is, When are they going to start charging you money to use LMStudio ?????? You got the best content as usual !!!!!!!!!!

[@horriblyblue6203]

Hello Matthew, does LMStudio supports AMD RX GPUs to power LLM? Because i can't figure it out, and it still uses my CPU only.

[@berer.]

Can you install Grok from the download provided by Elon Musk? Or do you always have to use their dl link?

[@shephusted2714]

you want to double down on this theme and ride it for the next few years or five - local, uncensored, real time, able to scan a wide variety of file formats, p2p sharing of models and training data plus no token limits or api calls - the way to go. maybe you can even delve into some open source llm mesh networking or something - i think that once the hw/sw matures and declines in price a bunch and things like cxl become common it is really going to light a fire under smb sector - lots of innovation, increased productivity and discovery.

[@alexlavertyau]

Does LLMStudio support any kind of image generation like MidJourney or ChatGPT plus ?

[@EM-yc8tv]

It's got multimodal image interpretation ability, with a Vision Adapter. For image generation, AFAIK LM Studio doesn't do that.... ComfyUI and Automatic1111 are what you'd be looking for to use all the Stable Diffusion models.

[@robottalks7312]

crew ai with claude 3 is it possible ? comparison with Devin

[@osamaa.h.altameemi5592]

does the server option in LM studio support concurrency?

[@matthew_berman]

Yes

[@osamaa.h.altameemi5592]

@@matthew_berman Now we a re officially in business. thx a ton for the fantastic videos.

[@user-bd8jb7ln5g]

I'm especially interested in the ability to generate multiple responses from the same model then selecting the best one. Can LM Studio do that at this time?

[@WylieWasp]

Hi Matthew just try to sign up for the newsletter something is broken just to let you know

[@duckpear2442]

2 questions: (1) do these models run offline (good for private data?); and (2) is LMStudio better than ollama?

[@MeinDeutschkurs]

Throttle: genius! LM-Studio got definitely improved. Gorgeous video!

[@dbookmine3046]

🎯 Key Takeaways for quick navigation: 00:00 *🌐 LM Studio simplifies running open-source large language models locally without coding knowledge, featuring a user-friendly interface and enhanced functionality.* 00:27 *🔄 The new update allows running multiple models simultaneously, making the process straightforward.* 00:54 *💻 LM Studio supports Mac, Windows, and Linux platforms, facilitating easy download and installation.* 01:08 *🔍 The software integrates with Hugging Face to search and download various large language models.* 01:35 *🤖 Features tools to predict if a model will fit on your computer, enhancing usability for newcomers.* 02:13 *📚 LM Studio includes built-in detailed documentation to assist users in understanding and utilizing different model quantizations.* 03:09 *📞 Introduces an AI chat tab, similar to ChatGPT, that uses local models, offering an engaging user interface.* 04:20 *🌱 Includes new features such as Branch conversation for testing and quality assurance.* 05:15 *⚡ Demonstrates significant speed improvements using built-in hardware acceleration like Apple's Metal.* 05:41 *🎮 The "Playground" section allows loading and managing multiple models simultaneously to enhance outputs.* 07:45 *💾 Introduces JSON mode to better integrate model outputs into applications, promoting ease of API utilization.* 10:17 *🚀 Shows how to run models simultaneously without sequential limitations, leveraging powerful computer capabilities.* 11:55 *🌐 Offers API capabilities to manage and execute model completions, extending usability for developers.* 13:14 *🤝 Highlights collaboration with other platforms like Autogen and potentially Crew AI, indicating LM Studio's flexibility in multi-agent environments.* Made with HARPA AI

[@gazorbpazorbian]

any idea of how I could share the models of ollama with this so I don't have duplicated models? also, for some reason, when you ask for a joke, it usually tells you the same joke... weird.. any idea of why?

[@starblaiz1986]

LMStudio has an option in its settings to specify a folder to download the models to. I haven't tried it myself, but I would assume simply pointing LMStudio to the same folder you've been downloading to for ollama would be all you need to do as the models are a standard format from Huggingface.

[@TheBlaser55]

Mathew, it would be great if there was a reference listing of all your videos so we could just pick a topic and find your videos that apply so we can watch them, again..... I think you did a video with agents before that did something simular, just wish they were easier to find

[@farorasyid1832]

hi. what common computer spec for this ? thanks

[@ricardocnn]

It also integrates with langchain and llamaindex

[@DaveEtchells]

Amazing to me how little RAM the models use. Time for me to get a maxed M3 Max MBP I guess, although gonna wait till after the May 7 event JIC. (I know very unlikely to have any MBP hardware impact, but I’m cautious 🙂)

[@Shari_Tejp]

can someone explain me importance of Ram VS Gpu vram? because for example on dolphin mixtral website the creator says you need at least 32 Gb of ram, and he's not mentioning nothing about GPU. and i taught the file size (for example 32 GB's) gets loded fully on your RAM, not on your gpu vram... im confused

[@JoaoWilliamRodriguesCardoso]

Would there be any app similar to LM studio for mobile phone? I would love to run these open source LLM on my phone.

[@u-save5989]

How to set it up on a server so I can sell access to chats of trained agents. Like GPTs in OpenAI shop. What GPU RAM CPU needed to use it per 5k daily users and if concurrent - how to calculate this and not crash server. Also - GROK, is it supported?