200k @ K : Q5_0 V: 4_1 (which is a bit of a sweet spot)
However, text-to-speech, speech-to-text, and non-code LLM use cases are so useful to have local, and don't require big hardware.
Having a universal reliable inference engine interface, I think, is the big unlock that needs to happen before app devs can ship these features.
Personal concrete use case: meeting recording app. This uses Parakeet + Qwen to create local transcriptions and post-cleanup, respectively.
Right now this app has to download and manage all these models, then bundle an inference engine to run them. It's a lot of code that probably should belong to the OS, or at least a standard interface.
While apps can offload some of this to llama.cpp or a similar process over http, that's another set of setup for the user to do before they can have a useful app.
Anyway, if you're getting started on a Mac, I'd suggest trying out oMLX (https://github.com/jundot/omlx) before messing with llama.cpp. In particular they have community benchmarks so you can see what kind of performance you're likely to get: https://omlx.ai/benchmarks. I wished each one had more configuration details though.
Certainly this is falsifiable easily by any of us doing it on a regular basis
> Qwen stuck in thought loops
This does happen when context is not managed effectively; creating plans, using subagents and compactions strategically resolves this
> creating plans, using subagents and compactions
Yes, these are all things that Claude Code does for you. However, for the thought loop issue, these are not the fixes. The canonical fix is to limit the number of thought tokens (llama.cpp's `--reasoning-budget`) or try to mess with the various penalty parameters. In any case, it's not a solved problem as far as I can tell.
How does that work? They have negative GPUs now!
don't get me wrong, the frontier models are leaps and bounds ahead of what qwen/kimikgemma are doing - but i don't need to drive a ferrari to the grocery store everytime either.
Personally I prefer the 35B MoE model, which is fast enough to be interactively useful, and capable, but I would probably use the 27B if I wanted to generate whole applications like that.
I am unconvinced that most "local" AI applications need anything much more powerful than the Gemma 4 12B model. Local agentic coding is a small niche, but there are plenty of ways a local model can help with development tasks.
I would really like to see a 12B or 16B Qwen 3.6.
I am currently playing with Ornith 1.0 in the MoE configuration, which is based on the 35B variant of Qwen 3.5; I am not sure if it is better than the 3.6 version.
Benchmarks say it is; my own silly tests either suggest otherwise or suggest that I have to talk to it a bit differently.
I really want to have a model that i can run locally on my 24gb m4 pro mbp for when i don't have internet to connect to my 3090 running the qwen, and i love how gemma 4 models 'feel', but i can't make them be competent. I am in the middle of finetuning both qwen3.5 9B and gemma 4 12B just to try and make those bridge closer to 27B for coding/agentic tasks (and am trying to ternarize and DQT 27B so that it fits in ~9gb pre-KV).
How do you run the gemma? What do you use it for (and in what harness), maybe llama.cpp and pi-mono just aren't for this model and that's what i'm doing wrong.
I've seen sites here and there but they feel like quick little toys that don't get updated, so they always suggest old models.
I haven't seen anyone make an argument they are as good as SotA (OpenAI, Anthropic). It's just they are approaching state where they are "as good" for some _limited_ set of use cases. Which will allow us to resolve 2 primary issues with these SotA models: privacy and vendor lock-in. Plus, they're very useful for education purposes, you get to explore what things looks like under the hood, play with various models, tools, maybe put something simple together yourself.
You get Macbook - great. You got gaming rig with a decent GPU - great (set it up as a dedicated server that you connect to through simple REST).
What exactly is wrong with any of that?
Is there any way to use MLX and GPU at the same time? Or does memory become a big problem?
TBH, I never understood Apple hyping these neural cores because I didn't think anyone actually uses them except maybe certain photo/video editing software.
If I can generate voice at the same time as video, that would be useful.
The neural cores aren't suitable for LLMs/transformers and isn't used in LLM inference. On the M5 and later chips, it comes with neural accelerators, aka Tensor Cores, which speed up the 'prefill' (i.e. processing your context window) part, but don't do anything for inference.
The MLX vs GGUF debate is mostly irrelevant. The GGUF pathways are optimised for apple silicon to the extent of practically identical performance to MLX. MLX is just one way of using Apple GPUs, it comes with many optimisations in the box, but they're not hard and they're no longer MLX-exclusive.
tweaking sampler might help
I'm running the NVFP4 alongside Gemma4 at the same quant on an OEM Spark
also i like that if i drop more sophisticated tools into my harness (e.g. any of the NLP/RAG-based search tools in place of grep/rg), the agent will actually reach for them and make progress faster; previous models have been reluctant to embrace new tools.
Lora if effective could be a great reason to run local models.
Qwen on the other hand got straight to work with astonishing competency on the same system.
From what I read llama3 needs beefier compute to reliably invoke tools, which I presume relates to it focussing more on simulating AGI rather than being a useful tool.
https://arena.ai/leaderboard/code/webdev/pareto?license=open...
https://arena.ai/leaderboard/text/pareto?license=open-source
I am still mostly tinkering/learning rather than spilling out code, and I feel quite slow on it. So it doesn't matter too much to me if it is really slow. More the journey than the destination if that makes sense. I'm stubborn.
I have tried the Gemma 4 12B model (Unsloth's QAT version) with search/browse tools in LM Studio and Unsloth Studio, when I am trying to understand a new thing.
Basically I get it to write introductory starter documentation for me to absorb, because my big personal problem, these days, is focussing enough to start a project and then digging in; I need the help.
I have found its limits on obscure packages (that it sometimes makes up) but before that it's a bit like stumbling on a blog post that happens to be really right for your particular need. Good enough to work through.
It's stuff I could ask Perplexity to do, or ChatGPT, to be fair, I just like LM Studio for this and have the inquisitiveness to want to run it locally.
In your case: I don't believe it's the quant. I'm sure it's the model — it has good coding knowledge but it's clearly not specialised. It might be good enough at writing Python/PHP/JavaScript at a novice level. It is also quite good on WordPress tooling and functions.
But I wouldn't bother with it for agentic coding if you've got experience elsewhere. Might be interesting to see what you can do with the 9B Ornith model?
Qwen 3.6 MoE in its Unsloth version is another matter. Impressive and I am trying to find ways to support my old brain doing what I've done before.
BUT DO NOT buy this MacBook if you plan on doing serious coding using local LLMs with it. The reason is simple: your fingers will burn and your head will explode from the noise.
Running any kind of sophisticated job on the very laptop you are using is just not viable. Sure you can use it in clamshell mode, but forget touching it while working with AI coding or agents.
If you want to run Qwen3.6 27B / 35B at its best, get a MacMini M4 with 64GB of RAM and put it in the basement - or at least a few meters from your desk. Connect to it over LAN or Tailscale. The MacMini will also cost you almost 1/3 of the MacBook Pro.
Thank me later.
So, just buy a mac mini and put it in the other room? ( Like everyone was doing in February? :)
I've been running coding agents on my laptop in yolo mode for the past half year or so (though mostly not local ones, laptop too slow!) and the way I'm doing that without terror is that I just gave them their own Linux user "agent". They're free to nuke their homedir /agent, and they can't touch (or even read) mine.
There's some slight ergonomics issues (I need to sudo into the user to do anything, but I set up an alias for it), sometimes I get issues with permissions or ownership (gave up on "sticky bits" and just made a function I can run once a day when it breaks).
There's enough hassle that I wish I just had a dedicated machine for it, and then I'd just give them root on it. (For giggles I gave claude root on a $3 VPS and that's going just fine...)
But yeah after months of trial and error I reinvented "just buy a mac mini" from first principles...
My hearing is not great, but I think I would have noticed the fan, and I have never heard it. In fact, I had to google to find out if it even has a fan.
Can confirm this works rather well, most things that integrate with LLMs, (agents, editors), support providing a remote (LAN) URL for Ollama, LM Studio etc.
But you do need a fast LAN connection, otherwise working with agents will be a pain.
Still, I don't agree. I think this machine is meant to use local models. You just have to wear pants if you want to keep it directly on your lap. I rarely use it that way anyway. I prefer it plugged into an external display and comfortably sitting on a laptop stand.
After about 1 minute the entire machine basically bricked and I had to hard reset :D
You would have to get a third party reseller/scalper or refurbished mac mini to get 64gb of ram ever since apple stopped selling it.
But it's also really easy to trip up. I fed it some of my Ars pieces and asked it to analyze themes and composition, and it got into a looping argument with me over how it was unable to analyze "my" writing because "the user cannot be the article author, the user is the user, the user did not write the article, the article author wrote the article." I was utterly unable to convince it that I was in fact me.
Qwen3.6-35B-A3B hums along at about 50GB of RAM used with --gpu-memory-utilization=0.42. I haven't tried Qwen3.6-27B (I'd likely grab Qwen3.6-27B-FP8, I think), but I'm curious to see if it makes much of a difference.
qwen3.6 35B A3B MLX 8bit -> 85-90 tok / sec! It is impressively fast and roughly 90% as good as 27B (in my opinion).
I can’t figure out when it makes sense to pay 10k up front for a quantized Llama 3.1 but it’s an interesting option
But man, I have never purchased a computer which is more expensive than a decent family car.
Some people will be happy to pay that premium for privacy, but at roughly 10X the cost of a MacBook Neo, that money could also buy a lot of credits on OpenRouter or frontier labs.
[0]: https://www.apple.com/shop/buy-mac/macbook-pro/14-inch-space...
The real test is whether or not it can work with your existing codebases. In my limited experiments Qwen 3.5 (maybe 3.6 is loads better) does OK on a Rust+React app, and less well on a C# monolith. Not to the point of being unusable but definitely poorly enough that I went back to Claude after 20 minutes. If I lost access to a cloud model and had to use Qwen instead I'd be visibly sad.
(I'm aware the price is, in absolute terms, more expensive where I live compared to the USA. That reinforces what I think, because anyone sane that would've bought one of those in another country would sell them as soon as they landed here and save that money.)
72.06 t/s. That's the full Qwen 3.6 27B model BF16, using MTP, running on Ollama. Yes I know I should bite the bullet and get vllm running on that box.
That was, also, at a 570 watt limit: I normally run a little less, but when I first tried this I actually forgot I had set the limit to 300 (it's a hot day, I figured why fight the A/C?), and at 300 watts the same question came back at 69.38 t/s. (The extra power matters more for compute bound things, the difference in generating LTX2.3 videos is considerably higher... but still not linear.)
QAT, MTP, 128k context.
I liked Qwen 3.6 27b too, it just seems that Gemma4 is a bit underrated.
Local development for who? How many of y'all are rocking 128GB of memory? Am I reading Apple's site correctly that it's a $10,000 laptop?
I’m not having it build whole features from scratch, though. I give it pretty explicit instructions closer to the class or function level, and it still saves me an immense amount of time, while I’m very connected to the code that’s written.
Definitely the sweet spot for me.
For 24GB VRAM cards (e.g. 4090) you can use Q6_K (22.5GB) or Q5_K_M (19.5GB) quants, possibly offloading some of the weights to RAM.
Unsloth recommends 18GB of RAM for Qwen3.6-27B (for their version of the model).
Sent from my 8gb M2 Mac mini.
On M5 128GB one can make use of the ram and use sparse MoE. For example, DeepSeek-V4-Flash will fit, served by DwarfStar (https://github.com/antirez/ds4). One will probably improve 2x the token/sec speed, given DS4F 13B activated params in the MoE are ~1/2 of the ~27B of the dense Qwen.
27B Of the Qwen fit even on a cheaper 24GB card, e.g. amd 7900xtx (<$1K?) or slightly dearer nvidia 3090 (with cuda). With ~900 GB/s bandwidth they will likely be ~50% faster than the M5 with 600 GB/s.
I ran those throu opus saking if it was good advice and was not impressed:
I read the actual qr_scanner.ino. Short answer: partially, but I'd push back on most of it. That review reads like generic ESP boilerplate advice written against an imagined version of your code — several of its "fixes" are already in your file, and its headline "critical" claim misreads what the code does. Going point by point:...
I just got a B70 with 32GB RAM for the equivalent of $1200 (incl. sales tax and import duties to my non-US location, so presumably it could be cheaper elsewhere). The memory bandwidth is 608 GB/s. For M5 Max (32-core GPU) it's 460 GB/s and for M5 Max (40-core GPU) it's 614 GB/s. A 3090 is still faster at ~900 GB/s but you're getting 32GB VRAM for a lot less than equivalent Nvidia cards. It's about 1/3 the bandwidth of a 5090 for 1/3 the cost, but with the same 32GB VRAM. If you're interested in being able to run bigger quants with some context and stay on a lower budget then it's an appealing trade off.
I'm still exploring using these local models so don't want to spend the equivalent of $5 000 - $10 000 just to test it out. I don't mind slightly slower perf to do some experimentation more affordably.
I actually got an B50 16GB (with meager 70w TDP!) first to test an Intel card with my stack - it worked easily with Ubuntu & Vulkan. I'd read a lot about hassles and people writing them off as unusable but it seems like these are often with SYCL which doesn't even seem to outperform vulkan and so why bother? (The B50 was just $370 inclusive tax and duties). Literally `apt install` the vulkan libraries and it worked with default xe driver in 26.04 and the vulkan build of llama.cpp. The SR-IOV PF/VF also just works with qemu/kvm, no tricks required. Since I got it fwupdmgr has updated the firmware twice so Intel is presumably actually trying to support these products.
It does about 30 tok/s which is enough for me. It's about half what the online models do, but it's enough.
I've heard their 9B models are also good, but they aren't much faster if you have the ram and a nice cpu.
These qwen3.6 models are the first ones I find can do much. GPT OSS was good, and Gemma4 is better. Gemma knows more facts, but qwen3.6 is smarter.
If a model runs fast enough for your use case and does exactly what you need it to, then you don't need a much slower model that might be more accurate. If you do anything more complicated, the dense models become more necessary and they are much more computationally heavy by comparison.
On your hardware an Unsloth quant of Gemma 4 26BA4B QAT would likely give you better results, but because it has 4B active parameters instead of Qwen's 3B active parameters, it will probably run slower.
Progress marches without mercy.
https://github.com/ikawrakow/ik_llama.cpp
Edit: it's gonna be slow if you're not using any VRAM. But it's possible. Software isn't going to speed that up anytime soon, it's just a hardware bandwidth limit.
Qwen 3.6 27B will run in full offload with a 4-bit quantisation in 64GB on an M1 Max. It is quite slow.
I don't know about 48GB but 64GB should be enough.
Jackrong has a few different ones available depending on what you're trying to do: https://huggingface.co/Jackrong
Was just trying to see how small I could go and get acceptable results, but yeah, larger Qwen 3.6 with MTP is going to be better. Cant wait to see how AI model (unsloth/local-llm/heretic/reaper/etc communities) are tweaking/engineering quality down into smaller models. Lots of new things coming out.
I do not have a crazy rig, a modest gaming one at that, but in trying to understand more about agents and their capabilities, I am SOL with my 16 GB of RAM and 8GB of VRAM. I can get most small, non tool calling models to perform well, but I've had major issues with anything over 9B doing anything more than reasoning (egregiously slow at higher parameter counts).
And so far, I cant get even Pi to extend itself or do any meaningful work with any of the models I currently can get to run.
For you, you could try gemma-4-26B-A4B
You should look at gemma-4-26B-A4B. 16+8=24gb and Q4 is about 16GB. Not much context left, but might run.
I very much appreciate the frank response, as it makes me feel less defeated at knowing my understanding of how it should work is not the full issue, hahaha
But certainly seems like we are a few years away from that, sadly.
Am I also screwed in being able to train my own small model or adjust another one with such a non-workhorse PC?
For anything else local, including writing some automation scripts and such, it works great.
>
> --jinja for tool calling support
Pretty sure this flag hasn't done anything for a while. It's enabled by default since ~November of last year
So for example I'd favour a used M1 Max over a used M2 Pro, at least based on my naïve understanding. Not quite sure where the balance changes.
There appear to be some hardware improvements with the M3 and up regarding the Apple Neural Engine which I'd hope would show up in MLX performance; I remember seeing some optimisations in image generation models that are only possible on later hardware.
The GPU cores are progressively better I believe, but the memory bandwidth is lower. Though perhaps the M4 can get closer to actually saturating said bandwidth.
(And I must reiterate that my understanding of this stuff is pretty naïve.)
A very useful resource for characteristics and comparative performance of all M variants, if anybody is interested, is https://github.com/ggml-org/llama.cpp/discussions/4167?sort=...
Its sister discussion for nvidia gpus is https://github.com/ggml-org/llama.cpp/discussions/15013
Note the drop in performance for the base (binned) m3 max version. You are better off with full m1 max than the binned m3 max, even price aside.
The issue I have with my m1 max is that with 64gb you cannot run really decent MoE models, ie the ones you can run like qwen 35B-A3B have only 3b active parameters and are much less capable than qwen 27b in my testing. So I end up running the 27b one, but it runs relatively slow (though still usable at 10-20 tok/s) and I would have been better off a used nvidia gpu setup for dense models. I assume 35B-A3B has its use cases, eg as subagents, just that I cannot find them. With a higher amount of ram I could probably run bigger MoE models which could be more comparable, though prefill would still be an issue (and prob a bigger one). The only hopeful thing is that there are performance hacks appearing (speculative decoding and prefill) that seem to start improving inference speed once getting implemented, so I am mildly hopeful.
(I must also iterate that my understanding is not very deep either)
It got rather tangled up when I tried it with one of my coding tests, which is a simple wordpress plugin, but I frustrate the model by asking it to write code for older PHP, break WP coding conventions and use a rather bespoke method for arranging code in objects. So it is sort of a hybrid of a green field and brown field task; a bit muddy.
It did not do as well as Qwen 3.6 35B, but the way it worked through its thoughts was interesting.
TBH I struggled to understand what DeepReinforce are doing that is materially different; the explanation of their training technique goes over my head at this point.
I don't know how much serious hands-free agentic coding I will ever do on my MacBook alone, but I do know that I would not have got so far into understanding this without tinkering with local models, llama.cpp, LM Studio, and LM Studio and all that.
I totally struggled to find the right frame of mind to explore any of this stuff without feeling defeated and bamboozled. Because it's just huge, exhausting, jargon-drenched, unknowable, and I am over the hill at fifty-plus.
Until, that is, I could poke around with setting it up on my own (secondhand) machine, watching the API calls, understanding some of the terminology. I didn't even buy the machine for that; it's just adequate to the task.
The Neo is too small to really get much benefit from this opportunity to make it more visceral and knowable.
Cloud models are (much) faster, they don't consume so much power/generate heat, they have much bigger (LLM) context, they're much more precise and they have a much wider (engineering) context of the given problem.
Except privacy and use cases that are blocked by cloud models (e.g. reverse engineering), local LLMs are currently an expensive toy.
When I try to program with a local LLM (I'm on a 32/128 GB system), I end up wasting time compared to a cloud LLM.
I don't recall any previous tech stack that was barfed onto the scene with so little background or reference material, going from zero to endless undefined jargon... and no primer in sight.
For people who demand an understanding of their tools, it's a lot of work. I recognize the value of "AI" in performing the tasks I'd have to do manually; for example, keeping the data structures of my front- and back-ends in sync in a project. But do I want to interrupt my development and take weeks off to digest all of these tools?
And if I do, I want to run the show and fully understand it. And like you, I think that's best done locally.
Agree having a powerful machine is really worth it in general for professionals, but strong disagree that running local LLMs has anything to do with it. It's hard enough as it is getting a good ROI on your time/money prompting/wrangling with frontier models. IMO leaning on the comparatively limited capabilities of local LLMs is best avoided in favor of keeping your own personal coding skills fresh and continuing to learn new ones.
I found LM studio to be a nice starting point. Frindlier and more featureful than Ollama and not as intimidating as llama.cpp (though you will want to use that eventually)
- opencode with it's webui
- deer-flow with it's research/powered front end
They both run websites so you don't have to baby sit them (eg, keep your mac open). I've build a pdf compressor over a few days by first having deer flow try and research the frameworks and pipeline. It stalls out because its not really a fluid programmer. Once it stalls out, I transferred it (manually for now) to opencode and it's refactoring it because it's just a collective bundle of sticks and it needs a lot of testing to tweak out the limited scop context. LLMs can't really hold large scopes (locally anyway, from what I've read from HN, it's possible with longer context).
It'll complete in a few days with maybe 3-4 hours of full attention interaction, but it's running 3x that without my attention. Obviously, if I paid more attention it'd run quicker, but since it's local, it's not pumping out large volumes of code, it's mostly looping over tests and capabilities as observed.
It's running Qwen3.6 35B MoE on a AMD 128GB strix halo. If I switched to the dense models, perhaps it'd be smarter, but the trade off seems to be much slower gen.
In theory you can also get 48GB of VRAM with, say, two 3090s, but it will take up a lot of space and generate a lot of heat compared to the Macbook Pro and GB10.
So like... $2000+ just for the used GPUs? Plus I assume it's considerably more effort to get it working.
The real sweet spot for Qwen 27B is getting it on something like a Dual 3090 system or some other config where it can blaze at 50-80 t/s and that costs well under 6K currently. It is a surprisingly capable model. Using something like GLM for orchestration, specs, task farming and then letting Qwen churn is relatively inexpensive.
Overall I recommend people try models of this class out using OpenCode and some for pay service to experiment with them and understand how they work. I find they are very useful.
Long term, I am convinced enough that if I wanted to use local models for any number of reasons I would be okay investing in a dual GPU box. The Mac is not fast enough for me and M5 Max is just too expensive relative to GPU linux box. Still, it is nice to have the models local ON the laptop and it is useful for what I care about locally.
Gemma 4 is the only model series at this parameter scale I've seen correctly answer some of these. One of the answers even made me re-evaluate what I thought the correct answer was, which I did not expect.
When I look at the Artificial Analysis numbers, I can see that some things about Qwen 3.6 look inflated as a result of either metrics that weren't measured yet for Gemma 4 31B, or for metrics that just aren't going to be relevant in a lot of the essential tasks. In a lot of the relevant metrics, Gemma 4 is either better or on par.
Then once it's all quantized all those benchmark results will be hurt, and Gemma 4 QAT has better quantized performance. I think it's more competitive unquantized than people give it credit for and way better quantized than people give it credit for.
Qwen 3.6 clearly isn't legitimately bad and maybe it's quite nice at fp16, but it was a disaster quantized in a 24GB scenario by comparison.
If you want to run unquantized, you definitely need 128GB.
Edit: it’s not just “data privacy”, when you are using Claude, you are shipping EVERYTHING to Anthropic. It’s crazy.
$5000 in US Treasuries (currently at 4.89%) yields $244.5/yr. That's more than enough to cover the annual Claude Pro subscription ($200/yr) which includes Claude Code with lots of Sonnet usage (far better than Qwen 3.6)
Qwen3.6-27B would be faster on a 3090 that costs around $1000-1200 though so I don't think it's a good counter-argument.
Op just happened to have that MacBook, but it doesn't mean it's necessary to run the model.
https://github.com/noonghunna/qwen36-27b-single-3090
Flies though (50-70tps is impressive for a model this smart)
I went through roughly the same process to get it working on my M2 Macbook Pro... at awful speeds of course, since models like this one are mostly bound by memory bandwidth.
I think you might be a little to into the stew here.
I haven't tried it with https://lemonade-server.ai/ yet but I just might give it a shot.
From what I understand, for a developer, $5000/month is maybe the high end, but $5000/year is fairly standard. (Is that accurate?) So if it pays back in 15 months, that's pretty decent. If it pays back in two months, that's spectacular.
Disclaimer: There's a 35% sale from Alibaba right now. And I'm not accounting for input tokens going faster than output tokens.
Do you know how much VRAM/unified is needed for the 27B model, which is generally regarded as better between the two compared in the article, is needed with little to no KLD loss and at 256k context?
Also, once you worked out how much memory is needed for that, maybe tell us how much a non-Apple system that you can run that (probably similarly or faster) would cost?
And when you have answered that, can you tell us how much privacy costs? Maybe also tell us how private OpenRouter is?
Edit: looking at other replies that are basically pointing out the same thing I did, I guess it's my wording. It's frustrating that people who misinform others in some nicely packaged ways or just simply uninformed get to keep doing that if they sound nice. Thanks.
Ryzen AI Max 395+ with 128GB of unified memory can be found around $3-4k.
But 27B isn't that large, either, especially if you are ok with the quantized models. So this laptop choice seems to more be a "because they had it" rather than "this is what's necessary for this particular workflow"
Not really germane to your comment but I hope I don’t sound old when I say I remember a time when spinning up a PoC was a week of work, and a statement like yours was pure science fiction.
If I start prompting away the core of a new project I lose interest in the entire thing almost straight away. I hate it. The next day I could care less about it. In fact it just makes me lazy, like a fat person who drives everywhere.
I love typing code and thinking for myself. Im going to continue to do that. I still dont know anyone who's shipped anything truly useful with this garbage tech, let alone with a local 30b param model. So much cope in these comments.
Spending 6k on hardware to run the worlds most mediocre model truly does make you an incredibly stupid person, so Im not really suprised by these comments of people saying these tiny models are helping them so much.
Its like a special needs kid all of sudden got the ability to code, of course they'd be impressed by basically all the code it produces.
This is an underrated consideration when evaluating the small models: The further you deviate from standard example code, the more their weaknesses show.
My experience is that Qwen3.6 produced some amazing results for a small model when I tried it with simple apps that are widely reproduced everywhere. If you want a React TODO app or to set up a little boilerplate app with shadcn and other popular tools, it will produce something that looks not too bad.
Then when I started straying outside of common tasks and into some of my more niche work, it would spin for hours and go in circles before finally producing some groan-inducing output that wasn't usable.
If you're looking for a model to help with simple refactoring or small tasks where you provide very explicit instructions for exactly what you want, but you don't want to do all of the typing yourself, they can do a lot of good work, though. But you're right that once you get into long context sessions involving topics off the beaten path, the weaknesses are very apparent.
The quantizations that are popular for making these models fit on smaller hardware make the problems worse. When you read it about online there is almost a consensus that 4-bit quants are lossless and that you can use q8_0/q8_0 kv cache quantization without any real loss, but in my experience with real projects there's a substantial degradation in long context performance with any of these quants.
Never go below an fp16 kv cache unless you've already tested it in advance with your model on a verified task that you know it can successfully complete. People should also test the difference using the exact same seed value so they can see how the tokens diverge. If you have memory constraints, sometimes you can still use an fp16 kv cache and use storage for an agentic buffer to work your task with mixed abstractions rather than having everything in memory.
For 4-bit weight quants, Gemma 4 31B QAT is where people should be looking instead of Qwen 3.6.
Modifying existing code is way easier if you don't expect it to be smart about it. Don't say "add X feature" and let it explore the codebase and build its own understanding. Point it at the relevant files and say "the goal is to add X feature to this code, follow Y guidelines". Now you've done the hardest part of making the decisions and it just has to follow instructions while coloring within the lines.
Is that not how you would work with any model, local or not? I wouldn't trust it to make the right decisions unattended. I just know the moment I look away it's going to do something utterly braindead.
1. Maybe you should tell us what those limited experiments are.
2. Maybe you should actually try 3.6 because it's huge difference in most cases. Don't forget to tell us quants and don't forget to tell us scope.
3. Maybe actually show us data compared to frontier models instead of this... vibe comment. Pretty tired of this kind of comments on HN that doesn't require logic or evidence. Just vibes. Like the pelican riding a bicycle crap that everyone has taken for granted but has no objective way of assessing goodness.
You get fewer tokens per second, but at some point the balance between quality and quantity makes the large model size worth the spend.
When you're spending this kind of money, you may as well treat yourself to a pretty screen and some decent speakers. Nothing the competition doesn't offer these days, but you get them for free with the car-priced RAM upgrade so why go for less.
Imagine its value if war broke out over Taiwan / Greater China, or really any of the dark scenarios with global connectivity or the truthiness of commercially available models. It is a very, very difficult piece of equipment to make at any other moment in history. I wish I could have purchased more. I saw the signs and price trends and out of stocks as they unfolded. No doubt others with the means are stockpiling.
Yes. Back in the my days at $faang in europe it was not uncommon to hear people getting 120-160 k€/year in compensation and we were “poor” compared to us engineers at the same faang (4-500 k$/year total compensation) with a bit of seniority…
Yes. Your people earn an order of magnitude less income than Americans.
Cloud models still feel ‘magic’, like you send a request off and get something back, like it’s something ‘special’. I used to joke that ChatGPT might be some kind of mechanical turk underneath.
Watching a model run local on your own machine hits different — you realise that yes, it IS just a computer program. Which for me actually makes me appreciate the leap we’ve made MORE, not less. From an information-theoretic point of view, LLMs really are something special.
The fact that they are just programs, that I’ve now experienced first-hand that they’re just programs, makes all those questions around consciousness and intelligence much more interesting.
I tried Ollama but I've settled on Unsloth Studio generally; once things really settle down I'll just run the llama-server UI, which is pretty nice.
A friend is tinkering with LLMs for amusement on a 16GB Raspberry Pi 5, and when I explained that llama.cpp now had a typical web chat interface he was so happy — it's amazing what the "table stakes" are now.
Have you tried Paseo?
I have opencode in a VM, and the paseo daemon running in the VM, and then the Paseo Mac app. Really nice.
(You can also use the Opencode GUI to frame a remote opencode web interface)
I'm gonna check out paseo, but am not looking forward to all the ram the agent needs + all the ram paseo needs
Nah, not really. It is a little annoying in terms of space and power, though. Not every case and motherboard can support cards that big.
The limited context is problematic. I’m not exactly sure what it’s got available but hermes was hit and miss on a prospecting job.
It does seem to be doing useful work but it’s not API call level quality
Even that isn't strictly necessary - you can get perfectly acceptable performance by splitting a model between multiple older 12 or 16 GB cards.
If the cost doubles, or 4x, which is seems to need to for them to go profitable, what then?
The 3090's TPD is 350W, but given that LLM's token generation isn't compute bound, people usually undervolt these cards to reduce power consumption. IIRC you can get as low as 200-250W without any degradation. Caveat these figures are without speculative decoding and at batch size =1.
I did find a few useful parameter settings I've already discovered using my single 3090 and ollama.
I'm just remarking that the LLMs overwhelm me with minutiae, especially as I'm working on code design. I frequently ask it to restate concisely, and that helps.
[edited to mention ollama as a nice alt]
I use my MBP essentially as my workstation, it's almost always plugged in. I have a MBA (M4, 24GB RAM) that I picked up for ~A$1500 or so, and that's an amazing daily driver. I don't do local LLM inference on that unit, I can just hit my own APIs (via LM Studio) on the MBP over Tailscale.
i've watched friends try that route; i've been through this before. taking a downgrade is never fun: if it's a thing you're likely to care about in the future, then sometimes it's better to place yourself in the right ecosystem early.
I’ve used Qwen 3.6 27B for many things at work, and I’m regularly able use it for reasonably scoped tasks.
I’m not saying these models are perfect.
But you are complaining about people on the extreme, while at the same shouting from the opposite extreme.
And I can't say that I won't switch to openrouter (even just for the same models) at some point.
But one of the things I have found about my own process learning is that some lessons only come to you when you make yourself available to them. And if that means doing things the difficult way, that is what you should do.
The rest of my life is ultra-frugal so I am relaxed about this.
I do realize the cloud is just someone else’s computer right? Power goes in, tokens and heat come out - just in another place
The interesting question is whether that gap will narrow, and if so, how much, and on what timescale.
The exact answer to this question is not knowable, but if you are the kind of person who comes to a site called "hacker news", and you think there is a nonzero chance that the answer is that yes, the gap will narrow and this won't always be an expensive toy, then now seems like a pretty great time to get in the game and start exploring the capabilities.
To me, "how do contemporary AI systems work and interact with contemporary hardware and how can I best take advantage of their capabilities?" is the set of skills that are worth learning at this moment.
What else is there? New / additional programming languages? New / additional database systems? frameworks? orchestrators? cloud provider / infra tooling? architectural patterns?
I dunno, all of this seems really boring and "been there done that" to me at this moment in time!
I needed to do this, this way, in my own time, to put my brain back together. It has worked for me, which is why I recommend it.
YMMV.
I have a pretty deep, maybe paranoid need to be confident I have an intrinsic understanding, and I have found in my life that lessons come to you when you make yourself open to learning.
So I need to build on top of what I know, taking as much of the hard way as I can bear to take at any one time — it has to be not quite difficult enough to put me off.
I can't really explain what I have learned this way that is different, but I feel it in a way that I wouldn't if I'd simply pushed a button.
For the same reason, I have a really basic 3D printer that I've set up myself, set up Klipper, configured how I want it, learned how to calibrate, all that. And now I can say that I feel I have an understanding of 3D printing. I could hold my head above water in a discussion with a real expert, maybe find work in an adjacent field where my insights would keep me grounded.
I can afford a really good printer that has all that set up, and more, has no problems. But I'd just be someone who has a 3D printer.
(Also who am I kidding about the existence of a printer with no problems)
I have colleagues that seem perfectly content to delegate too much to the agents, and it saddens me. It feels like there will be swaths of engineers that didn't train some of the critical thinking skills that I take for granted.
I certainly see it in slack discourse around anything more complicated than a feature implementation. Maybe I'm just cynical. Time will tell, I suppose.
I mean one of the things I use a local LLM for, because I can, is to generate starter documentation. But I ask it to — I want it to give me overviews, plans, all that. It can make something bespoke for me.
I guess I could also ask it to do the work. But where do you draw the line?
The universal labour-saving device is the great provocation of the next 100 years I think, and both Star Trek and Wall-E have grappled with it.
And that's how skills die.
In the open model space an insane amount of effort goes into getting more powerful models to run with the same or less RAM. For example in the diffusion world many things that could not be run on easily under 24GB of VRAM actually run much better today with much less VRAM than they did a few years ago. You can do many things today with 8-16GB of VRAM that would not have been possible. At the same time the most advanced open models, like LTX 2.3 for video gen, still seem to respect 24GB of VRAM as the upper bound.
Similarly the standard "big" but localish open model for LLMs back in the day was Llama 3 70B, this was both a much worse and much larger model than Qwen 3.6 27B
So in two different spaces I've witnessed the "RAM required to run the best" decreasing or at least remaining stable, while the performance being achieved in both areas is astounding (LTX 2.3 is faster, better and more capable than the Wan 2.2 model that held popularity before it).
The biggest thing to watch out for is not just RAM/VRAM but memory bandwidth. You can try to "future proof" yourself with lots of RAM, but if it's 400 GB/S you're still constrained to smaller models.
I'm thinking of getting a SoC machine with 128GB RAM but the bandwidth is limited to 256 GBps. Would you even consider such a machine a decent investment, or should I wait for the newer gen of chips? Thanks!
The same can be said about operating system memory requirements. I am sure Linux and Windows kernel developers can confirm. Yet 30 years ago Solaris used to run comfortably in 16 MB of RAM, today you need 512 times that to run Linux.
at 128GB, you can find almost it's entire context for Qwen3.6 35B MoE.
Again, I think you have too much faith in extrapolation. It's like you got a baby at 0 months, then measured it at 12 months and expect it to be a giant.
But if this is the case, as you say, it seems like a good opportunity to build a more welcoming set of entry points into this!
(Very reminiscent of 3D printing, where you get a lot of very trivial advice poorly applied, which is an analogy I've now made several times.)
Several of the youtubers are pretty helpful, though; I watched half a dozen things and absorbed the broad pattern and then went for it.
Also I got a lot out of reading HN comments, which is why I am here; tucked away in the corners of these discussions are people who can help. Over time I hope I am one.
... but, the models that WILL run on 128GB (or 64GB or even 32GB) models today are a huge improvement on the best models that would run in the same amount of memory six months ago.
> GLM-5.2 class models already need 1TB+ of RAM.
If you find three finds that also have a 128GB MacBook, you can chain them together (the MacBooks, not your friends) and make it work.
You could also run GLM-5.2 on a single MacBook if you stream the active parameters from disk, but even with speculative decoding, you'd probably only get in the order of 1 token per second, so this is not really practical for most applications.
I don't know if it has changed my mind about a career change but as I am sure you can understand, I no longer feel like I am running away defeated.
My very best wishes to you :-)
Having spent a good weekend learning how to perform latent-steering through playing with pytorch and a local Gemma4 model, there is no way I could have groked any of that in the the way I did without hands on time.
This is on an M3 Max 36GB I've had for a couple of years. No further outlay needed.
(sarcasm, btw)
For example (and relevant to AI) I can generate electricity on my roof at $0.20-25/kWh, batteries included. In California the electric utility can’t offer it cheaper than $0.30-0.50/kWh. Therefore at scale, electricity is actually more expensive.
There are many such examples.
That's never the point of keeping local alternatives though.
For me this dates all the way back to installing Slackware 1.0 (0.99pl12!) on an offline 486SX rather than just using the internet-connected workstations in the lab.
Here, I already had a Mac that was powerful enough to run a local LLM, so now I do, because I can.
The reason I delegate so much of local LLM installation and administration to Claude Code is simply because there's no point learning practical things that will work completely differently in a couple of years, or in memorizing procedures that I'll forget long before I need to perform them again.
No longer having to sweat all the details is a Good Thing, not a Bad Thing.
But I think if you want to really learn to ride well, understand horses well, there might be some benefit in learning how to shoe a horse. At some level it should never only be someone else's job.
Maybe a more apt analogy would be a skill like making fire without a lighter.
Maybe my biggest problem with the world of agentic AI, and the reason I am putting myself through learning it the way I am, is that the need to know the "why" of everything is so fundamental to me, that I don't know if there is any point to me without it.
So this is really the only way I know how to proceed.
And we happen to be discussing this on a forum where the type of people who will be the specialists for the system were discussing are likely to gather.
I'd be surprised if in my casual discussions out in the real world, I were to run into a lot of people who care exactly how all this works, to the extent that they want to invest significant money into hardware that allows them to run things themselves and dig into what's actually going on. But I'm not at all surprised to come across such people here! (Indeed, it would be very disappointing if I didn't!)
For example, you need to know it uses gasoline (or diesel), it requires oil changes every certain amount of time, break pad replacement, etc.
You also probably need to know that you can't operate cars over a certain amount of water, that you need a driver's license, stopping at red lights, etc.
Sure, you might not need to be a mechanic, but that's far from not understanding how a car works, which to me sounds similar to knowing how to shoe a horse, which is different than being a horse vet.
Right now, there is way more scale in centralized AI than there is at the edge. But that could flip. I'd still probably put the probability that it will under 50%. But I'd also put it above zero!