Comment by c7b
10 hours ago
Gotta say, I've lost all interest in cloud-based AI products. Too many cool features and workflows that I was once excited about that I can't or don't use anymore for a variety of reasons (price hikes, subjectively nerfed, disappeared altogether, replaced,...) for me to even remember. It's tiring.
I've set up a small rig, mostly settled on Qwen3.6 and I'm slowly adding features myself. It probably can't compete with Claude. I don't even know, I've stopped checking. It's providing a ton of value to me as is, and it only keeps getting better. All it takes is to realize that it doesn't actually matter if the grass is (maybe even objectively) greener somewhere else. Feels so good to know that it won't change under my feet. I've got this amazing, highly extensible tool, and it's mine.
I'm really happy this is one of the top comments here, I am fully local as well.
Just wanted to leave a note for folks who might not have the memory to run a big 32gb model - I just found out there are some pruned models that have really good performance and If I had a smaller machine I might try this pruned unsloth Q4 quant of GLM 4.7 flash that sits at 14gb: https://huggingface.co/unsloth/GLM-4.7-Flash-REAP-23B-A3B-GG...
I usually use LM Studio for this type of thing but unsloth has their own studio type app that might be even better suited for these quants.
I used GLM 4.7 flash as my main model for months and it was an incredibly tenacious model and very very fast - I think on restricted hardware, this could be a great choice.
Qwen3.6-35B-A3B-UD-Q4_K_M runs at about 11 tokens/second on my poor old 1060. Absolutely nuts how far we've come
I tried running any model on my 1070 and it instantly crashes my old tower, probably time to get off windows and run linux on it.
Understated how much of a boon for Linux that AI development has been.
There isn’t any benefit to running a windows machine.
1 reply →
Sounds like a hardware issue, though NVIDIA driver issues can't be ruled out, they're much rarer these days
Mind sharing your llama.cpp settings for that?
Using this llama.cpp fork https://github.com/TheTom/llama-cpp-turboquant and mostly copying from this video https://www.youtube.com/watch?v=8F_5pdcD3HY
Haven't had much time to test it other than asking a few questions & changing some HTML in cline so it might be thick as a brick for all I know, but still worth trying
1 reply →
I often feel like we're nowadays mostly pushing AI developments in the ways of finetuning differences. Like how new editions of Claude are tuned for agentic coding which might even be detrimental if you're using it for non-agentic coding. Or how Fable 5 in fact do look great but at a huge cost for inference and a high likelihood of post-launch nerfs or limit/price revisions. How Gemini 3.5 has more liberal limits but on the other hand underperforms a bit.
It's like we're mostly treading mud at this point. New editions are released, a version number increases, but I have to wonder if all steps are forward or they're more just tuned differently with similar actual perf per dollar as when this year began.
Most in fact seem to be happening to me with small models. Like your Qwen. Or Gemma 4 31B which is kinda magic especially when considering multilingual abilities. So yes, in that sense I can see "development" probably as we refine data sets and training methods but I see it less on the big hulking beasts with daily limits (unless you turn it up to 11 like Fable).
Edit: As I posted this, I saw a "before and after" comparison for Fable and the reintroduced version is seeing a catastrophic drop in BridgeBench performance as they're still mucking with the model. Go figure... https://x.com/Hesamation/status/2072692225100612032
Same here, been happy throwing Qwen3.6 on my old MBP - no it's not as fast as Claude which I use at work, but it works well enough locally and I don't have to worry about credits or shit like the rug getting pulled under me in terms of capabilities.
This sounds very appealing. What size Mac mini would I need for that?
Personally, I would always max out the RAM you can fit into your budget. You might get lower bandwidth (= slower generation) than you do on a Mac if you choose a Strix Halo or DGX Spark, but there are always new tweaks being discovered to speed things up. That being said, with 32GB you should be able to fit an ok quant of 35B-A3B or 27B with some context, with 64GB you should be golden.
i have issues on a m5/64g with 35b-a3b (mlx) it eventually hits a memory cap around 52gb... but i'm pretty happy with `Qwen3.6-27B-Claude-Opus-Reasoning-Distilled-mlx-8Bit`
1 reply →
A PC with an nvidia card with 16gb vram works just fine for Qwen MoE models, and these have worked great as a daily driver for me.
A 4-bit quantization of either Qwen 3.6 27b or Gemma 4 31b will run on a 32GB Mac with a decent-sized, but not full-sized, context. 64GB gets you the full ~256k context and you don't need to quantize your KV cache (though 8-bit quantization of KV may be worth it for performance). The 4-bit QAT version of Gemma 4 has practically identical performance to the full size version or the 8-bit version in most benchmarks and my tests, so there's no reason to run anything else. The 4-bit Qwen is a little bit lossy, as it hasn't gotten the QAT treatment, but not catastrophically lossy. A 6-bit dynamic quantization would be better for that model, but it's ~25GB on disk, and you'll need more than 32GB to run it with a big context.
I wrote up how I run local LLMs, with numbers and a focus on running Qwen 3.6 and Gemma 4. I prefer Gemma 4 31b, even though the general consensus is that Qwen 3.6 is better for code, and it is better on most coding focused benchmarks...it doesn't seem to be for my use cases, Gemma feels smarter. And, with QAT, you get more smarts in less memory, so it's fast and runs on more hardware.
https://swelljoe.com/post/how-i-run-local-llms/
Currently, the sweet spot for self-hosted models is either Qwen 3.6 or Gemma 4, and those top out at 31B (Gemma) and 35B (for Qwen, but you want the dense Qwen 3.6 27B if you can run it as reasonable speed...the dense models are much smarter), so for now, a system with 64GB or 128GB is going to be running the same models. Going to a bigger model doesn't get you better performance because there aren't any better models that are a little bigger. I wish there was a ~70B or even ~120B MoE in the Qwen 3.6 or Gemma 4 families, as I've got a Strix Halo running a model that leaves a lot of memory on the table (and it's not very fast, to boot...an MoE would be faster, and hopefully smarter if it's a much bigger model, like double or triple sized).
In short, right now, 64GB is all you need for the best models you can self-host on anything short of five-figure machines, but, I wouldn't buy any hardware right now, if you can wait a while. Tokens from DeepSeek are so cheap, you can wait out the memory shortage and get access to models you could never host locally. And, OpenRouter always has free models in preview or just because that you can use lightly, as they're rate-limited (but your self-hosted models are going to be rate-limited, too, because a Mac Mini can't run models very fast). Google AI Studio has the Gemma 4 models for free too, also rate/usage limited.
Good summary blog: https://maloyan.xyz/blog/running-qwen-locally-mac-mini-m4
> That's not hypothetical — it's a real measurement on the base model Mac Mini.
Hmmm
I am curious if you implicitly assumed they are Macs or if that's what you are looking for specifically?
I assumed the 27B dense model would be preferable to a MoE model, and that it wouldn’t fit into a consumer graphics card, which leaves the Macs.
Then I assumed for cost and battery/heat reasons that a Mini would be better than a laptop.
7 replies →
People want to make it seem like you need to always use the latest and greatest frontier models to be taken seriously as a developer.
You really don’t need them. After a certain point, bigger models give diminishing returns. If you can get 80% of the productivity gain with a free local model, use the local model. It will still be way faster than doing everything by hand, but you also don’t have to pay for tokens to a cloud provider and the tools won’t be ripped away from you on a whim.
This is the new attitude enlightened people should adopt. Reject the arms race.
The biggest appeal of the frontier models is for those trying to get autonomous agentic systems running that do real work with minimal human input. I went down a rabbit hole trying that with frontier models, and after a lot of initial promise it ended up actually slowing me down.
We've all been through that no? In the beginning you can do a ton of stuff without reading code. But the LLMs miss all the good abstractions, they just push and push unmaintainable code until at some point you start having more bugs and then you NEED that LLM to fix the codebase you don't understand anymore.
There are guardrails you can and must add to protect your team if you take the vibe approach: a good type system, a good database with clearly written business model and a good data model to drive your business. Make it loud and clear when something breaks with your tooling.
But... I'd definitely not vibe everything after a certain point. Reading and fixing code is also a lot of fun.
4 replies →
> People want to make it seem like you need to always use the latest and greatest frontier models to be taken seriously as a developer.
Except you kinda do. Try getting a job today without mentioning Claude experience. In another year it'll probably be something else. Saying you like to use Copilot today makes one seem elderly.
Not saying you need frontier models on a technical basis, but for career PR you probably do.
I never got into any of the AI models because it was clear local first was going to be more valueable, if they were to replace coding tasks.
I tried out a few models and ended up going with either Qwen3-Coder-Next (no think, just do) and Qwen3.6-35B (thinking, w/llamacpp token budget). Created a customized prompt that works fairly well to around ~60k tokens and then is a toss up on whether it's poisoned itself or I've directly steered it into the wrong. When it's clear that's happened, if it's important to continue, ask it to write a doc then start fresh.
I don't kno whow any one cold have witnessed the last 2 decades of American VC funded tech startups and tell themselves, "you know, this will be a reliable technolgy with no hidden problems".
Even a sober technical evaluation is just two steps:
1. You're proposing to build a app on a non-deterministic model.
2. That model is hosted behind a non-deterministic system (model alignment, model guardrails, system context subterfuge, cost/token pricing)
---
So you want to build your app and you think you're going to kep up with both #1 and #2?
Cool! Anything you want to share? I haven't looked much into my system prompt yet, do you have any tips?
We live in a non-deterministic world. Anything "deterministic" in it is a castle built on quicksand.
LLMs are, as far as the nastiness of the Real World goes, really fucking benign. Future models outperform past models, both in open weight land and at the big frontier labs. Performance per $ only ever goes up. That's just nice.
> We live in a non-deterministic world. Anything "deterministic" in it is a castle built on quicksand.
Except the Enterprise, and a lot of what people want compute for, is built on deterministic systems or processes. I'm not saying the non-deterministic nature of LLMs isn't useful. However I've worked with a lot of organizations on SOAR projects, for example. When you can weave the deterministic and non-deterministic together you get a relatively efficient system. A workflow that will stay on the rails and will come to a conclusion as expected. And the "as expected" part is critical in these types of systems. The reality of, using SOAR as an example, is also that most enterprise would be much better served by fast SLMs. Parse an email and validate if it's SPAM / Phishing or read a chunk of firewall logs and look for outliers / indications for escalation - those things can get messy in a deterministic system because of potentially unstructured data.
I don't believe it's either / or. And I believe that LLMs just aren't efficient, fast or reliable in the sense that deterministic are. It seems, at least to me, a better together story.
1 reply →
YES, but you seem to not understand that having two non-deterministic layers is incompatible. #1 is fine: it has random issue and you build around those random issues; those issues don't change unless you change them.
#2 is not fine; that non-determinism you do not control, have no insight into, etc.
I'm saying sure, give me #1 if it means I can build a harness around it and smooth over the edges. But I'm not taking #1 and #2. There's zero reasonable way to manae two non-deterministic systems.
Qwen is the Alibaba distilled Anthropic Claude model
So piracy on an by piracy trained ai model..
Piracy? Lol.
Alibaba didn't steal Opus weights, they used opus output to train their model.
If this is piracy, then so is reverse engineering efforts powering a bunch of Linux drivers.
1 reply →
Well, Anthropic got paid for it, unlike the sources that they used...
I'm not sure what you're trying to say. Is that a good or a bad thing? Model distillation is presumably part of the reason why Qwen is so good, yes. As a consumer, that's a good thing I would say. It's a natural counterbalance to the monopolistic tendencies of other tech segments.
If you have ethical concerns, model distillation feels like an arbitrary line to draw. Why is the first type of piracy ok, the second not? You should restrict yourself to ethical open source models. Which is btw where I genuinely hope the future of local models is going to lie. Open weights is not enough, we need fully open source models to be sustainable. Even for simple things like updating the knowledge cutoff. How we are going to distribute the training effort will be an interesting problem where I don't see an obvious solution yet. Maybe the blockchain/federated learning people can suggest something. Or university consortia, or some public sector solutions. Or something really boring - I for one would absolutely be willing to pay for DRM-free weights of an open source model (even if I could pirate them for free).
2 replies →
Same here, I’ve removed my credit card from Copilot and won’t be renewing
What features/workflows have you added?
Web search, MTP (speeds up generation), uncensored models. Lots more things on my bucket list (eg various things related to image generation).
Not gonna lie, if you're coming from ChatGPT/Claude Code, you'll mostly be adding back features you've taken for granted, or solving problems you wouldn't have had. But sometimes you do get some extra utility, like uncensored models, which have become my go-to. Not because I'm doing anything saucy, but I hated how I'd become trained to pre-emptivly self-censor my prompts. The guardrails in open weights models are no less strong than in proprietary ones, subjectively even a bit stronger in Qwen. But luckily there's an entire sub-discipline of model ablation. Another advantage would be better control over image generation (although I can't attest to that, yet).