Lets set aside speed. The results are ALWAYS terrible compared to CLaude Code.
Sure I only have a 12GB GPU - is that what it is? If I had a 32GB CPU would the experience be as amazing as Claude Code is?
Why is my experience of local LLMs so average?
You can maybe replicate Opus 4.5 to 4.7 if you have some moderately expensive hardware. And that is definitely good enough for many people.
The smallest model that can even pretend to be a decent coding model is Qwen3.8 27B, which you realistically need 32GB of VRAM to run without breaking or limiting it in some really noticeable way. This gets you something like a slightly less capable Opus 4.5 with a much better work ethic.
They seem to be happy with their small usages.
For everything I do, i see huge differences in Quality and Speed.
Even when I try an open model on a 4090, the PC gets warm/hot very fast, I struggle with context length.
> They seem to be happy with their small usages.
Oh yes, we're terribly inferior people. Just living on scraps from the masters' tables.
I do plenty of experiments of things i always wanted to do and they just don't work with local models.
Finetuning of some ML models for a specific usecase, some 4k demo idea i had, small browser games to show others game mechanic ideas i had for ages, analysis tools etc.
Even just playing around with some Ghidra eats context like nothing.
I write 5-20 paragraphs of things I want and it goes away and just builds it.
Something like this (just an example):
"Build an application with a webserver in rust. It has 4 stages, first one is an input module... second one is refinement ... third one is an image detection with current fitting ml model ... last one is finetuning."
Then i would describe every basic page with a few more paragraphs.
This would work for half an hour or an hour, consume 100-400k of context tokens and then I would test it, write down 10-50 small sentences (every sentence either fixing a bug, or making it better) which tehn takes again like 20 minutes and i'm nearly done with a complex enough tool providing real value and actually working.
A big kimi, which is hard to run, might be getting close to this but we do not talk about running it on 1-2 5090.
The other thing is, thats more like a practical experience, even if you use them through Open Router (for testing and co), the initial results of something like Opus 5.5 are ALWAYS better. More details, better quality, better understanding of usability of the tool.
And yes Sonnet 5.5 is crazy good
> In a work tree build a feature where each column cog action menu has a "Column statistics" item - clicking that item opens a modal dialog which displays statistics about that column - for text columns that's most common values, stats about value length - for numeric columns it's more interesting than that. Let's do a bit of percentile analysis in there too. Under the hood it works by firing off SQL queries using SQLite window functions and suchlike. For the moment let's have that option only available if SQL queries are allowed against the database, and the SQL queries are built in JavaScript.
You can see a screenshot of what Opus 5.5 built here: https://github.com/simonw/datasette/pull/3038
Having currently a very tech-intensive chat running for more than one week, this is where the frontier models really shine: Its no more 2024 or so, where they lost context on large discussions, you can do really great work with the frontier models - the local models will be a thing in a few years, if RAM & CPU developed further, this will happen: in 2035 like, you will be able to run "Claude-like experience" on your iPhone.
- 2023 using Cursor
- 2024 using Cursor with Agent mode
- 2025 using Claude Code
- 2026 using Codex with Astra 6
I've try every new model when it comes out from both OpenAI and Anthropic. I cannot justify leaving Codex and Astra 6. The token burn on Anthropic is always astronomical for me. Codex just works, and I see no reason to move away from it.
There is, for sure, a limiting factor in the x86 architecture, but I think the RTX Spark will change that a bit, and maybe AMD's subsequent Ryzen AI/Strix Halo chips.
In any case, 128-256GB unified memory systems running MOE models is the "cheap" option, RTX6000 is the step above it.
I'm not familiar with these. How much does one cost, and from what store / what product name?
I am digging in on Pi (in lovely Paseo and exploring Pendant when in VS Code).
What would I find better or worse in Kilo?
The newer/larger models are good though, not as good as Claude, but not every task requires a model of that level. Maybe yours do, but for simple coding tasks, summarising text, etc a small local model runs just fine, and is a lot cheaper than Claude
You can try that model (and many others) using a cloud host first, and then only if you like it do you need to buy something to run it.
RTX 5070 Ti and RTX 5080 both have 16 GB.
RTX 5090 has 32 GB.
Good luck finding any of them for reasonable prices, though. I paid $3,000 for my 5090 early last year. I winced when I bought it then. Today, it's like $7,000, and I wouldn't even consider it at that price.
But I'm mostly a gamer that also does local AI as a hobby. If the RTX 5090 was $7,000, I'd still be on my RTX 3080 and still just not started doing local AI and have to turn down the graphics settings in Cyberpunk 2077.
You can surely build an LLM server with an R9700 and 32GB of mainboard RAM for less than the 64GB M5 Max Apple Studio, which is probably the next most competitive model up. It'll be better than a Strix Halo but less flexible.
- 4090: Going for $7500
- 64GB M5 Max: Starts at $3,800
- R9700: $2700
Baseline of $20 or $100/month Claude or OpenAI subscriptions. Alternative posed of running models locally. A guy says he did it and results are lackluster on a 12GB video card. A number of people post that that card is insufficient; use 32GB+ cards. I look up prices of 32GB cards, and they are all very expensive. So from a practical perspective, people are suggesting you spend thousands of dollars on a video card to get results that are almost as good as the subscription plan they are on. It doesn't add up, unless one of my assumptions is incorrect.
Right now you could extensively use an M5 Max for a year and sell it for almost exactly the same amount at the end. Maybe even more, at the current rate. And it can do other things for you: host your VMs, edit high def video, whatever. Even if it stops being useful to serve LLMs, it might be useful to run as a server for other tasks for a whole decade. It will be a fast, supported Mac for seven years. For the R9700, maybe gaming?
Depends what you want, doesn't it? Or are we all expected to conform?
For me personally, right now I have a secondhand M1 Max macbook pro, and that is my local AI machine, because my needs are exploratory and I want to run it here. I want to learn by comparing models. I want to learn how to run them and advise others how to (local inference is not a zero-sized market).
My next step up will be to take the apps I'm looking at building that run against these small models, and host those apps online with some sort of cloud inference — openrouter, digitalocean, cloudflare — for those same models.
If I ever find I want to vaguely describe some desktop application before bed and come down the next morning to an aggravating slop version of it written in Rust, I'll consider Claude.
For a local non-laptop, I've looked at the M5 Max and the R9700. I've also looked at the M5 Pro, which in practice is about as fast at token generation as my M1 Max (because the M1 Max's memory bandwidth is strictly theoretical), and smokes it at prefill. That alone might be enough for a "leave it running" mini AI box, and it will be a nice desktop machine. I just don't need it enough yet.
As an aside there's this thing — somewhat interesting, though constrained:
Uses the 5090 mobile chip, which is pretty good speedwise (about the same as the R9700) but 24GB. Small, tidy, on-prem. I think they have missed the boat on pricing, but if they can ship in volume, maybe not.
3D games give me motion sickness so I don't have powerful PCs anymore. In fact right now I don't think I have a desktop/server PC that can power up at all, and that alone makes me wonder about building an R9700 machine.
Having a 4090 at home running agents all day is not a $20pcm level of usage, so take the $100 plan. Looking at the S-1 filing, today they are not even close to breakeven. It is reasonable to assume that will become a $300-400 pcm plan.
The launch MSRP of a 4090 was $1600. There has been significant inflation since then, so say $2000. Eventually supply will meet demand and you'll be able to buy one for that price again, or lower due to the economies of scale they must have now.
So, in 2-3 years time once it has all settled, you pay off your 4090 by saving 5-6 months of Claude usage. Maybe the local model isn't good enough for complete replacement, but all it needs to do is drop you to the $20 plan to be worth it. I think this is a large part of why investors are getting jittery now.
However even if you had a private data center you wouldn't be able to beat Opus 5.5, Astra or Gemini 4. The open weights models are a tier below these three.
Good enough? Quite probably, unless you're in the "coding is solved, specify loops and walk away" frame of mind.
32GB (or say a 48GB-unified RAM unified machine) gets you Qwen 3.8 27B at a good adaptive Q4, maybe Q5. Muse Glimmer at Q4 and Q5, etc.
These models can code well enough in an agentic way, I find, and the chat-to-assistant scenario is just fine in Unsloth Desktop or LM Studio.
A slightly lighter harness helps; a lighter system prompt is entirely possible, especially for Muse Glimmer, which is post-trained so effectively for agentic helper tasks that it barely seems to need a system prompt at all except for a "plan" mode.
Tool use is all fine, all that.
which is where the entire software industry is going
But will you want to do that when your tokens are metered at non-subsidised prices, and yet there is absolutely no clarity (ever!) about token efficiency, how the model decides to spend your money, how it might ever decide to stop wasting your money and get to the point, etc.?
The specify-loops-and-walk-away model absolutely points to local/on-premises/private cloud AI.
Because otherwise you're letting Anthropic or OpenAI decide how to spend your money. And they are going to be biased towards spending more of it.
I don't think Qwen 3.8 is quite there yet with the loops stuff. Definitely seems to be better at deeper and deeper one-shot puzzles.
"SemiAnalysis, a research and analysis company specializing in the semiconductor and AI industries, found that a $200/month subscription plan is far more generous than previously thought. According to its calculations, while it's widely believed that a $200/month plan maxes out at roughly $2,000/month in tokens (assuming API pricing), the actual numbers are closer to $14,000 for ChatGPT and $8,000 for Claude. "
https://cybernews.com/ai-news/chatgpt-claude-usage-costs-ope...
This will change in the future.
Then you would go try it and be back using windows to game a week later.
Everything considered, linux is good and has decent capabilities including gaming and whatever you can throw at it but only do oranges to oranges comparison. Games and apps built for each os works fine on the respective os, well atleast it is true for mac and linux. Windows is very unreliable.
Your experience is average because it _is_ supposed to be average. The people using local models tradeoff taking the moral/technological high ground vs actually using something high quality to get a job done quickly and with high quality, and largely autonomously. They are ok with subpar output because they will handhold the model more, or the (subpar relative to the frontier) outputs are sufficient for their toy use case while being cheap and accessible, with the added benefit of being able to run it on their private data without sending it to a 3rd party.
Almost no one at any job where software is the main economic output is using a local model to code. If you actually require intelligence to get a hard job done, you should just use a frontier model.
Or, you know, your own intelligence.
What I do not understand as an engineer is how our community has gone from believing in open source for controlling destiny, security, etc., from avoiding unnecessary corporate capture (think of all the rants about Swift!), from believing in heterogeneity and portability, to just saying, look, if you're not using the product of one of two loss-making AI unicorns to do all of your work, you're just being a contrarian weirdo who is doing it for performative reasons.
I guess if that is your definition of contrarian — that everyone else is doing it so why aren't we? — then I am.
But in my experience it's not so different to "why aren't you on social media?" — which was an argument from insecurity and the desperate need to fit in.
I personally believe in independent thinking. I have things I want to learn about how all this actually works, I have questions about my own life and future, and I'm not going to simply chuck them away because you or anyone else thinks it's just being contrarian and we shouldn't question the direction everyone is taking. So I am learning the tech with a stack I control, while continuing to do the job I do.
You can do what you want. You're probably right, and I'll be quite happily wrong and continue in my wrongness even though, based on how you characterise it, this kind of independent thinking appears to unsettle you.
How much effort did you spend to make the local setup work well?
Local models are not plug and play, you need to adjust settings to make them behave intelligently.
On my side one model looking super dumb right after install ended up being 'agentic' after using the right chat template, fixing how tool calls are setup, fine-tuning plenty of params values, and adjusting the whole setup to my own hardware. Claude.ai (free) was great at helping me sort that all out.
If that were really the case, basically nobody would spend so much money using them.
With my DGX Spark + 5080 + 3060 setup, I've tried all kinds of approaches, including quantization, acceleration, fine-tuning, post-training, and everything else.
It still can't compare to Opus 5.5 max in terms of speed, capability, and stability.
And when I calculate all the money I've invested, directly subscribing to the service would actually be cheaper, more efficient, and more stable.
Open source is great, but don't glorify it.
The architecture of the code I develop has never been simple. If you're someone who doesn't even work on things that need to be deployed to something like k8s, then basically, open-source models are indeed enough for you.
Found your issue.