Just ordered a $12k mac studio w/ 512GB of integrated RAM.
Can't wait for it to arrive and crank up LM Studio. It's literally the first install. I'm going to download it with safari.
LM Studio is newish, and it's not a perfect interface yet, but it's fantastic at what it does which is bring local LLMs to the masses w/o them having to know much.
Exo is this radically cool tool that automatically clusters all hosts on your network running Exo and uses their combined GPUs for increased throughput.
Like HPC environments, you are going to need ultra fast interconnects, but it's just IP based.
RTX is nice, but it's memory limited and requires to have a full desktop machine to run it in. I'd take slower inference (as long as it's not less than 15tk/s) for more memory any day!
If the primary use case is input heavy, which is true of agentic tools, there’s a world where partial GPU offload with many channels of DDR5 system RAM leads to an overall better experience. A good GPU will process input many times faster, and with good RAM you might end up with decent output speed still. Seems like that would come in close to $12k?
And there would be no competition for models that do fit entirely inside that VRAM, for example Qwen3 32B.
RTX Pro 6000 can't do DeepSeek R1 671B Q4, you'd need 5-6 of them, which makes it way more expensive. Moreover, MacStudio will do it at 150W whereas Pro 6000 would start at 1500W.
I'm using it on MacBook Air M1 / 8 GB RAM with Qwen3-4B to generate summaries and tags for my vibe-coded Bloomberg Terminal-style RSS reader :-) It works fine (the laptop gets hot and slow, but fine).
Probably should just use llama.cpp server/ollama and not waste a gig of memory on Electron, but I like GUIs.
8 GB of RAM with local LLMs in general is iffy: a 8-bit quantized Qwen3-4B is 4.2GB on disk and likely more in memory. 16 GB is usually the minimum to be able to run decent models without compromising on heavy quantization.
I'd love to host my own LLMs but I keep getting held back from the quality and affordability of Cloud LLMs. Why go local unless there's private data involved?
There are some use cases I use LLMs for where I don't care a lot about the data being private (although that's a plus) but I don't want to pay XXX€ for classifying some data and I particularly don't want to worry about having to pay that again if I want to redo it with some changes.
Using local LLMs for this I don't worry about the price at all, I can leave it doing three tries per "task" without tripling the cost if I wanted to.
It's true that there is an upfront cost but way easier to get over that hump than on-demand/per-token costs, at least for me.
Same. For 'sovereignty ' reasons I eventually will move to local processing, but for now in development/prototyping the gap with hosted LLM's seems too wide.
I did this a month ago and don't regret it one bit. I had a long laundry list of ML "stuff" I wanted to play with or questions to answer. There's no world in which I'm paying by the request, or token, or whatever, for hacking on fun projects. Keeping an eye on the meter is the opposite of having fun and I have absolutely nowhere I can put a loud, hot GPU (that probably has "gamer" lighting no less) in my fam's small apartment.
Right on. I also have a laundry list of ML things I want to do starting with fine tuning models.
I don't mind paying for models to do things like code. I like to move really fast when I'm coding. But for other things, I just didn't want to spend a week or two coming up on the hardware needed to build a GPU system. You can just order a big GPU box, but it's going to cost you astronomically right now. Building a system with 4-5 PCIE 5.0 x16 slots, enough power, enough pcie lanes... It's a lot to learn. You can't go on PC part picker and just hunt a motherboard with 6 double slots.
This is a machine to let me do some things with local models. My first goal is to run some quantized version of the new V3 model and try to use it for coding tasks.
I expect it will be slow for sure, but I just want to know what it's capable of.
I genuinely cannot wrap my head around spending this much money on hardware that is dramatically inferior to hardware that costs half the price. MacOS is not even great anymore, they stopped improving their UX like a decade ago.
If the rumors about splitting CPU/GPU in new Macs are true, your MacStudio will be the last one capable of running DeepSeek R1 671B Q4. It looks like Apple had an accidental winner that will go away with the end of unified RAM.
Not OP, but with LM Studio I get a chat interface out-of-the-box for local models, while with openwebui I'd need to configure it to point to an OpenAI API-compatible server (like LM Studio). It can also help determine which models will work well with your hardware.
LM Studio isn't FOSS though.
I did enjoy hooking up OpenWebUI to Firefox's experimental AI Chatbot. (browser.ml.chat.hideLocalhost to false, browser.ml.chat.provider to localhost:${openwebui-port})
i recently tried openwebui but it was so painful to get it to run with local model.
that "first run experience" of lm studio is pretty fire in comparison. can't really talk about actually working with it though, still waiting for the 8GB download
Yup, I'm spoiled by Claude 3.7 Sonnet right now. I had to stop using opus for plan mode in my Agent because it is just so expensive. I'm using Gemini 2.5 pro for that now.
LM Studio has quickly become the best way to run local LLMs on an Apple Silicon Mac: no offense to vllm/ollama and other terminal-based approaches, but LLMs have many levers for tweaking output and sometimes you need a UI to manage it. Now that LM Studio supports MLX models, it's one of the most efficient too.
I'm not bullish on MCP, but at the least this approach gives a good way to experiment with it for free.
This isn't true. You can `ollama run {model}`, `/set parameter num_ctx {ctx}` and then `/save`. Recommended to `/save {model}:{ctx}` to persist on model update
I just wish they did some facelifting of UI. Right now is too colorfull for me and many different shades of similar colors. I wish they copy some color pallet from google ai studio or from trae or pycharm.
MCP terminology is already super confusing, but this seems to just introduce "MCP Host" randomly in a way that makes no sense to me at all.
> "MCP Host": applications (like LM Studio or Claude Desktop) that can connect to MCP servers, and make their resources available to models.
I think everyone else is calling this an "MCP Client", so I'm not sure why they would want to call themselves a host - makes it sound like they are hosting MCP servers (definitely something that people are doing, even though often the server is run on the same machine as the client), when in fact they are just a client? Or am I confused?
I think host is a bad term for it though as it makes more intuitive sense for the host to host the server and the client to connect to it, especially for remote MCP servers which are probably going to become the default way of using them.
The initial experience with LMStudio and MCP doesn't seem to be great, I think their docs could do with a happy path demo for newcomers.
Upon installing the first model offered is google/gemma-3-12b - which in fairness is pretty decent compared to others.
It's not obvious how to show the right sidebar they're talking about, it's the flask icon which turns into a collapse icon when you click it.
I set the MCP up with playwright, asked it to read the top headline from HN and it got stuck on an infinite loop of navigating to Hacker News, but doing nothing with the output.
I wanted to try it out with a few other models, but figuring out how to download new models isn't obvious either, it turned out to be the search icon. Anyway other models didn't fare much better either, some outright ignored the tools despite having the capacity for 'tool use'.
Gemma3 models can follow instructions but were not trained to call tools, which is the backbone of MCP support. You would likely have a better experience with models from the Qwen3 family.
Others mentioned qwen3, but which works fine with HN stories for me, but the comments still trip it up and it'll start thinking the comments are part of the original question after a while.
I also tried the recent deepseek 8b distill, but it was much worse for tool calling than qwen3 8b.
Great to see more local AI tools supporting MCP! Recently I've also added MCP support to recurse.chat. When running locally (LLaMA.cpp and Ollama) it still needs to catch up in terms of tool calling capabilities (for example tool call accuracy / parallel tool calls) compared to the well known providers but it's starting to get pretty usable.
What models are you using on LM Studio for what task and with how much memory?
I have a 48GB macbook pro and Gemma3 (one of the abliterated ones) fits my non-code use case perfectly (generating crime stories which the reader tries to guess the killer).
Readit News
Can't wait for it to arrive and crank up LM Studio. It's literally the first install. I'm going to download it with safari.
LM Studio is newish, and it's not a perfect interface yet, but it's fantastic at what it does which is bring local LLMs to the masses w/o them having to know much.
There is another project that people should be aware of: https://github.com/exo-explore/exo
Exo is this radically cool tool that automatically clusters all hosts on your network running Exo and uses their combined GPUs for increased throughput.
Like HPC environments, you are going to need ultra fast interconnects, but it's just IP based.
Get the RTX Pro 6000 for 8.5k with double the bandwidth. It will be way better
The whole point of spending that much money for them is to run massive models, like the full R1, which the Pro 6000 cant
If the primary use case is input heavy, which is true of agentic tools, there’s a world where partial GPU offload with many channels of DDR5 system RAM leads to an overall better experience. A good GPU will process input many times faster, and with good RAM you might end up with decent output speed still. Seems like that would come in close to $12k?
And there would be no competition for models that do fit entirely inside that VRAM, for example Qwen3 32B.
Probably should just use llama.cpp server/ollama and not waste a gig of memory on Electron, but I like GUIs.
Using local LLMs for this I don't worry about the price at all, I can leave it doing three tries per "task" without tripling the cost if I wanted to.
It's true that there is an upfront cost but way easier to get over that hump than on-demand/per-token costs, at least for me.
Oof you were NOT joking
I don't mind paying for models to do things like code. I like to move really fast when I'm coding. But for other things, I just didn't want to spend a week or two coming up on the hardware needed to build a GPU system. You can just order a big GPU box, but it's going to cost you astronomically right now. Building a system with 4-5 PCIE 5.0 x16 slots, enough power, enough pcie lanes... It's a lot to learn. You can't go on PC part picker and just hunt a motherboard with 6 double slots.
This is a machine to let me do some things with local models. My first goal is to run some quantized version of the new V3 model and try to use it for coding tasks.
I expect it will be slow for sure, but I just want to know what it's capable of.
LM Studio isn't FOSS though.
I did enjoy hooking up OpenWebUI to Firefox's experimental AI Chatbot. (browser.ml.chat.hideLocalhost to false, browser.ml.chat.provider to localhost:${openwebui-port})
Deleted Comment
I'm interested in using models for code generation, but I'm not expecting much in that regard.
I'm planning to attempt fine tuning open source models on certain tool sets, especially MCP tools.
I haven’t been using it much. All it has on it is LM Studio, Ollama, and Stats.app.
> Can't wait for it to arrive and crank up LM Studio. It's literally the first install. I'm going to download it with safari.
lol, yup. same.
I'm considering ordering one of these today: https://www.newegg.com/p/N82E16816139451?Item=N82E1681613945...
It looks like it will hold 5 GPUs with a single slot open for infiniband
Then local models might be lower quality, but it won't be slow! :)
Dead Comment
Dead Comment
I have one running locally with this config:
1. CodeRunner: https://github.com/BandarLabs/coderunner (I am one of the authors)I'm not bullish on MCP, but at the least this approach gives a good way to experiment with it for free.
You gotta help me out. What do you see holding it back?
> "MCP Host": applications (like LM Studio or Claude Desktop) that can connect to MCP servers, and make their resources available to models.
I think everyone else is calling this an "MCP Client", so I'm not sure why they would want to call themselves a host - makes it sound like they are hosting MCP servers (definitely something that people are doing, even though often the server is run on the same machine as the client), when in fact they are just a client? Or am I confused?
Some more discussion on the confusion here https://github.com/modelcontextprotocol/modelcontextprotocol... where they acknowledge that most people call it a client and that that's ok unless the distinction is important.
I think host is a bad term for it though as it makes more intuitive sense for the host to host the server and the client to connect to it, especially for remote MCP servers which are probably going to become the default way of using them.
https://modelcontextprotocol.io/specification/2025-03-26/arc...
Upon installing the first model offered is google/gemma-3-12b - which in fairness is pretty decent compared to others.
It's not obvious how to show the right sidebar they're talking about, it's the flask icon which turns into a collapse icon when you click it.
I set the MCP up with playwright, asked it to read the top headline from HN and it got stuck on an infinite loop of navigating to Hacker News, but doing nothing with the output.
I wanted to try it out with a few other models, but figuring out how to download new models isn't obvious either, it turned out to be the search icon. Anyway other models didn't fare much better either, some outright ignored the tools despite having the capacity for 'tool use'.
I also tried the recent deepseek 8b distill, but it was much worse for tool calling than qwen3 8b.
I'd love to learn more about your MCP implementation. Wanna chat?
Nice to have a local option, especially for some prompts.
I have a 48GB macbook pro and Gemma3 (one of the abliterated ones) fits my non-code use case perfectly (generating crime stories which the reader tries to guess the killer).
For code, I still call Google to use Gemini.
Deleted Comment
Deleted Comment