I appreciate the author's modesty but the flip-flopping was a little confusing. If I'm not mistaken, the conclusion is that by "self-hosting" you save money in all cases, but you cripple performance in scenarios where you need to squeeze out the kind of quality that requires hardware that's impractical to cobble together at home or within a laptop.
I am still toying with the notion of assembling an LLM tower with a few old GPUs but I don't use LLMs enough at the moment to justify it.
In my experience the latest models (Opus 4.5, GPT 5.2) Are _just_ starting to keep up with the problems I'm throwing at them, and I really wish they did a better job, so I think we're still 1-2 years away from local models not wasting developer time outside of CRUD web apps.
> I realized I looked at this more from the angle of a hobbiest paying for these coding tools. Someone doing little side projects—not someone in a production setting. I did this because I see a lot of people signing up for $100/mo or $200/mo coding subscriptions for personal projects when they likely don’t need to.
Are people really doing that?
If that's you, know that you can get a LONG way on the $20/month plans from OpenAI and Anthropic. The OpenAI one in particular is a great deal, because Codex is charged a whole lot lower than Claude.
The time to cough up $100 or $200/month is when you've exhausted your $20/month quota and you are frustrated at getting cut off. At that point you should be able to make a responsible decision by yourself.
The limits for the $20/month plan can be reached in 10-20 minutes when having it explore large codebases with directed. It’s also easy to blow right through the quota if you’re not managing content well (waiting until it fills up and then auto-compacting, or even using /compact frequently instead of /clear or the equivalent in different tools).
For most of my work I only need the LLM to perform a structured search of the codebase or to refactor something faster than I can type, so the $20/month plan is fine for me.
But for someone trying to get the LLM to write code for them, I could see the $20/month plans being exhausted very quickly. My experience with trying “vibecoding” style app development, even with highly detailed design documents and even providing test case expected output, has felt like lighting tokens on fire at a phenomenal rate. If I don’t interrupt every couple of commands and point out some mistake or wrong direction it can spin seemingly for hours trying to deal with one little problem after another. This is less obvious when doing something basic like a simple React app, but becomes extremely obvious once you deviate from material that’s represented a lot in training materials.
> The time to cough up $100 or $200/month is when you've exhausted your $20/month quota and you are frustrated at getting cut off. At that point you should be able to make a responsible decision by yourself.
leo dicaprio snapping gif
These kinds of articles should focus on use case because mileage may vary depending on maturity of idea, testing and host of other factors.
If the app, service, or whatever is unproven, that's a sunk cost on macbook vs. 4 weeks to validate an idea which is a pretty long time.
If you're a hobbyist doing a side project, I'd start with Google and use anti-gravity, then only move to OpenAI when the project gets too complex for Gemini to handle things.
This story talks about MLX and Ollama but doesn't mention LM Studio - https://lmstudio.ai/
LM Studio can run both MLX and GGUF models but does so from an Ollama style (but more full-featured) macOS GUI. They also have a very actively maintained model catalog at https://lmstudio.ai/models
Cline + RooCode and VSCode already works really well with local models like qwen3-coder or even the latest gpt-oss. It is not as plug-and-play as Claude but it gets you to a point where you only have to do the last 5% of the work
"This particular [80B] model is what I’m using with 128GB of RAM". The author then goes on to breezily suggest you try the 4B model instead of you only have 8GB of RAM. With no discussion of exactly what a hit in quality you'll be taking doing that.
I'm curious what the mental calculus was that a $5k laptop would competitively benchmark against SOTA models for the next 5 years was.
Somewhat comically, the author seems to have made it about 2 days. Out of 1,825. I think the real story is the folly of fixating your eyes on shiny new hardware and searching for justifications. I'm too ashamed to admit how many times I've done that dance...
Local models are purely for fun, hobby, and extreme privacy paranoia. If you really want privacy beyond a ToS guarantee, just lease a server (I know they can still be spying on that, but it's a threshold.)
> Local models are purely for fun, hobby, and extreme privacy paranoia
I always find it funny when the same people who were adamant that GPT-4 was game-changer level of intelligence are now dismissing local models that are both way more competent and much faster than GPT-4 was.
Moon lander computers were also game changers. Does not mean I should be impressed by the compute of a 30 year old calcualator that is 100x more powerful/efficient in 2025 when we have stuff a few orders of magnitude better.
For simple compute, its usefulness curve is a log scale. 10x faster may only be 2x more useful. For LLMs (and human intelligence) its more quadratic, if not inverse log (140IQ human can do maths that you cannot do with 2x 70IQ humans. And I know, IQ is not a good/real metric, but you get the point)
There is a cultural component to privacy, and insulting people who want a technical assurance that their queries are under their own control is a way to work against privacy.
I wouldn't run local models on the development PC. Instead run them on a box in another room or another location. Less fan noise and it won't influence the performance of the pc you're working on.
Latency is not an issue at all for LLMs, even a few hundred ms won't matter.
It doesn't make a lot of sense to me, except when working offline while traveling.
I don’t think I’ve ever read an article where the reason I knew the author was completely wrong about all of their assumptions was that they admitted it themselves and left the bad assumptions in the article.
The above paragraph is meant to be a compliment.
But justifying it based on keeping his Mac for five years is crazy. At the rate things are moving, coding models are going to get so much better in a year, the gap is going to widen.
Also in the case of his father where he is working for a company that must use a self hosted model or any other company that needed it, would a $10K Mac Studio with 512GB RAM be worth it? What about two Mac Studios connected over Thunderbolt using the newly released support in macOS 26?
He actually addressed your point by pointing out that, in his view, the models this machine can run today are the worst it’ll ever be - he thinks, and my own experience over the last year, is that local models improve and at a pace which is CLOSING the gap to commercial models.
If privacy is your top priority, then sure spend a few grand on hardware and run everything locally.
Personally, I run a few local models (around 30B params is the ceiling on my hardware at 8k context), and I still keep a $200 ChatGPT subscription cause I'm not spending $5-6k just to run models like K2 or GLM-4.6 (they’re usable, but clearly behind OpenAI, Claude, or Gemini for my workflow)
I was got excited about aescoder-4b (model that specialize in web design only) after its DesignArena benchmarks, but it falls apart on large codebases and is mediocre at Tailwind
That said, I think there’s real potential in small, highly specialized models like 4B model trained only for FastAPI, Tailwind or a single framework. Until that actually exists and works well, I’m sticking with remote services.
Not worth it yet. I run a 6000 black for image and video generation, but local coding models just aren't on the same level as the closed ones.
I grabbed Gemini for $10/month during Black Friday, GPT for $15, and Claude for $20. Comes out to $45 total, and I never hit the limits since I toggle between the different models. Plus it has the benefit of not dumping too much money into one provider or hyper focusing on one model.
That said, as soon as an open weight model gets to the level of the closed ones we have now, I'll switch to local inference in a heartbeat.
I’ve been using Qwen3 Coder 30b quantized down to IQ3_XSS to fit in < 16gb vram. Blazing fast 200+ tokens per second on a 4080. I don’t ask anything complicated, but one off scripts to do something I’d normally have to do manually by hand or take an hour to write the script myself? Absolutely.
These are no more than a few dozen lines I can easily eyeball and verify with confidence- that’s done in under 60 seconds and leaves Claude code with plenty of quota for significant tasks.
Are people really so naive to think that the price/quality of proprietary models is going to stay the same forever? I would guess sometime in the next 2-3 years all of the major AI companies are going to increase the price/enshittify their models to the point where running local models is really going to be worth it.
This is not really a guide to local coding models which is kinda disappointing. Would have been interested in a review of all the cutting edge open weight models in various applications.
My takeaway is that clock is ticking on Claude, Codex et al's AI monopoly. If a local setup can do 90% of what Claude can do today, what do things look like in 5 years?
I just got a RTX 5090, so I thought I'd see what all the fuss was about these AI coding tools. I've previously copy pasted back and forth from Claude but never used the instruct models.
So I fired up Cline with gpt-oss-120b, asked it to tell me what a specific function does, and proceeded to watch it run `cat README.md` over and over again.
I'm sure it's better with other the Qwen Coder models, but it was a pretty funny first look.
Can anyone give any tips for getting something that runs fairly fast under ollama? It doesn't have to be very intelligent.
When I tried gpt-oss and qwen using ollama on an M2 Mac the main problem was that they were extremely slow. But I did have a need for a free local model.
79 comments
[ 4.0 ms ] story [ 78.0 ms ] threadI am still toying with the notion of assembling an LLM tower with a few old GPUs but I don't use LLMs enough at the moment to justify it.
Are people really doing that?
If that's you, know that you can get a LONG way on the $20/month plans from OpenAI and Anthropic. The OpenAI one in particular is a great deal, because Codex is charged a whole lot lower than Claude.
The time to cough up $100 or $200/month is when you've exhausted your $20/month quota and you are frustrated at getting cut off. At that point you should be able to make a responsible decision by yourself.
For most of my work I only need the LLM to perform a structured search of the codebase or to refactor something faster than I can type, so the $20/month plan is fine for me.
But for someone trying to get the LLM to write code for them, I could see the $20/month plans being exhausted very quickly. My experience with trying “vibecoding” style app development, even with highly detailed design documents and even providing test case expected output, has felt like lighting tokens on fire at a phenomenal rate. If I don’t interrupt every couple of commands and point out some mistake or wrong direction it can spin seemingly for hours trying to deal with one little problem after another. This is less obvious when doing something basic like a simple React app, but becomes extremely obvious once you deviate from material that’s represented a lot in training materials.
leo dicaprio snapping gif
These kinds of articles should focus on use case because mileage may vary depending on maturity of idea, testing and host of other factors.
If the app, service, or whatever is unproven, that's a sunk cost on macbook vs. 4 weeks to validate an idea which is a pretty long time.
If the idea is sound then run it on macbook :)
in my experience cursor is nicer to work with the openai/anthropic cli tools
Sure am. Capacity to finish personal projects has tripled for a mere $200/month. Would purchase again.
From what my team tells me, it's not a great deal since it's so far behind Claude in capabilities and IDE integration.
LM Studio can run both MLX and GGUF models but does so from an Ollama style (but more full-featured) macOS GUI. They also have a very actively maintained model catalog at https://lmstudio.ai/models
Somewhat comically, the author seems to have made it about 2 days. Out of 1,825. I think the real story is the folly of fixating your eyes on shiny new hardware and searching for justifications. I'm too ashamed to admit how many times I've done that dance...
Local models are purely for fun, hobby, and extreme privacy paranoia. If you really want privacy beyond a ToS guarantee, just lease a server (I know they can still be spying on that, but it's a threshold.)
I always find it funny when the same people who were adamant that GPT-4 was game-changer level of intelligence are now dismissing local models that are both way more competent and much faster than GPT-4 was.
For simple compute, its usefulness curve is a log scale. 10x faster may only be 2x more useful. For LLMs (and human intelligence) its more quadratic, if not inverse log (140IQ human can do maths that you cannot do with 2x 70IQ humans. And I know, IQ is not a good/real metric, but you get the point)
Latency is not an issue at all for LLMs, even a few hundred ms won't matter.
It doesn't make a lot of sense to me, except when working offline while traveling.
The above paragraph is meant to be a compliment.
But justifying it based on keeping his Mac for five years is crazy. At the rate things are moving, coding models are going to get so much better in a year, the gap is going to widen.
Also in the case of his father where he is working for a company that must use a self hosted model or any other company that needed it, would a $10K Mac Studio with 512GB RAM be worth it? What about two Mac Studios connected over Thunderbolt using the newly released support in macOS 26?
https://news.ycombinator.com/item?id=46248644
Personally, I run a few local models (around 30B params is the ceiling on my hardware at 8k context), and I still keep a $200 ChatGPT subscription cause I'm not spending $5-6k just to run models like K2 or GLM-4.6 (they’re usable, but clearly behind OpenAI, Claude, or Gemini for my workflow)
I was got excited about aescoder-4b (model that specialize in web design only) after its DesignArena benchmarks, but it falls apart on large codebases and is mediocre at Tailwind
That said, I think there’s real potential in small, highly specialized models like 4B model trained only for FastAPI, Tailwind or a single framework. Until that actually exists and works well, I’m sticking with remote services.
I grabbed Gemini for $10/month during Black Friday, GPT for $15, and Claude for $20. Comes out to $45 total, and I never hit the limits since I toggle between the different models. Plus it has the benefit of not dumping too much money into one provider or hyper focusing on one model.
That said, as soon as an open weight model gets to the level of the closed ones we have now, I'll switch to local inference in a heartbeat.
These are no more than a few dozen lines I can easily eyeball and verify with confidence- that’s done in under 60 seconds and leaves Claude code with plenty of quota for significant tasks.
So I fired up Cline with gpt-oss-120b, asked it to tell me what a specific function does, and proceeded to watch it run `cat README.md` over and over again.
I'm sure it's better with other the Qwen Coder models, but it was a pretty funny first look.
When I tried gpt-oss and qwen using ollama on an M2 Mac the main problem was that they were extremely slow. But I did have a need for a free local model.