Is there any hope for people that cant even run 27B parameters, Qwen3.6 or otherwise? Are there any quantized models that do well with tool calling at smaller parameter sizes?
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.
The article is based on running Qwen 3.6 on a 128GB MacBook Pro. For reference, a 128GB MBP currently starts at $6699 USD [0]
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.
The maths there is pretty undeniable, but it is not where I'd make the split. Having a machine that can run some modest local LLMs, like the Gemma 4 12B, is really worth it.
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.
I just got Claude to download and install all the models and servers and agents and prepare all the launch scripts for me... no need to learn, just ask it to do it for you
> I totally struggled to find the right frame of mind to explore any of this stuff without feeling defeated and bamboozled.
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)
The model they reference can be easily run with 24gb+ of VRAM, and there are other similar models capable of running easily on 16gb of VRAM. It's not like 128gb is a requirement here.
You can also run Qwen 3.6 27B dense model on DGX Spark with comparable performance [1][2] for about $4000 (Asus Ascent GX10 is $3999 at various retailers).
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.
Just putting it out there: I run Qwen 3.6 on my M1 Mac Studio with 64gb. It's quantized and all that, but I agree with TFA: it's the sweet spot for local development right now.
I’m running the same model on a 48GB MBP with a q4 quant and it’s pretty decent. You definitely don’t 128GB. That’s the scale for 70B models at q8 or something.
How much does one of those cost in the US? Here in Brazil, your notebook is worth as much as a used Honda Fit, which seems absolutely insane. For comparison, the ThinkPad I'm currently running cost me 1/20 of how much this MBP costs here, leaving me with over $8.000 to spend with LLM inference (if I actually spent money with that).
I purchased mine for approximately $4400 AUD before the price hikes. That unit is now ~$5100 AUD.
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 run Qwen 3.6 on my Framework Desktop 128GB, and it's very performant. I know Framework has had to raise the price since I preordered mine, but they're still well under half the cost of that Macbook.
Absolutely for the average developer the token speed is just going to be too slow for it to be workable. I think we’re looking at 2028 when memory becomes cheaper again and they’ll be a lot more people using local models.
My experience working in the open model space pretty deeply (both LLMs and diffusion models) for years now is that it is not quite as simple as that.
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 eyars. 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 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.
It can't run the latest models today - GLM-5.2 class models already need 1TB+ of RAM.
... 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.
How many credits would it buy? How long would it take to use them up? What's the payback period?
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.
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.
> maybe tell us how much a non-Apple system that you can run that (probably similarly or faster) would cost?
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"
That's my point. You can run Qwen3.6 27B with MTP and whatever else you want to bolt onto it at 256k context for much less than even a Ryzen AI Max 395+ with 128GB would cost. Even unquantized you don't need 128 GB so given your comment and the downvotes maybe I didn't word my original comment properly for this?
I've been working with local models for the past year. There's so many possibilities, but I don't think coding is one. Coding requires so many layers beyond inference; I spent so much time trying to replicate what Claude Code does end to end locally. Understanding all the layers and keeping up with the advancements feels like a slog. Even this article messes up and misunderstands what some of the settings are doing. Qwen in particular seems to work at first, then often gets stuck in thought loops when used for actual work.
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.
None of the examples reflect 'real work', at least not what I'd consider real work. Being able to nail a zero-shot greenfield project is relatively easy even for a small model. There's not much context to build up and it can fall back to similar examples in the training data easily. So long as you're not asking it to invent something wholly new it'll probably manage.
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.
In my experience, even with basic project concepts the small models struggle to spin up greenfield stuff. There's just too many decisions to be made and they're not good at that.
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.
> In my limited experiments Qwen 3.5 (maybe 3.6 is loads better)
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.
27-30B in general seems to be the level where you actually start having decent models. I just wish consumer hardware hadn't stagnated so much that we can't easily go higher than that, and that even running those requires a $5k machine now.
I feel like I'm going insane seeing people buy these 128gb MBP for thousands of dollars to run models that are objectively much worse than SOTA and spending so much more. The amount spent on a 128gb M5 MAX can buy you a damned new car here. What the hell am I missing? Are developers in other countries living in such different worlds?
(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.)
My experience also aligns with this. I'm running gemma4 31B on a 4090 through llm.cpp with unsloth models.
I also run Qwen 3.6. Qwen is good for thinking and planning as it is faster, but Gemma4's generated code is much higher quality in the first try (Rust, C++ and C#). so it needs less revisions to be at a level I'm comfortable for merging.
Nice. I flip flop between Qwen 3.5 9B Q6_M and Gemma4 12B Q4_K_M on a 4080 Super. They run at about the same speed and I can have them review each other's plan or diffs. For smaller projects I find them very capable, and I can step up to a better quant for slightly more challenging work.
This is what I run on an M5 MacBook Air 32GB. Works great.
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.
Think commercial. My company invested in a local rig since privacy is important to our customers and sometimes I want to use these models on private data.
I have been having pretty good success with Qwen 3.5 9B for "nontrivial but not challenging work all things considered" -- it runs great on my 24gb unified memory m4 pro MacBook Pro. What do the baseline specs look like Mac-wise for getting this model to run? Am I looking at a 96gb? 128? 256?
Thanks! I was thinking of doing the 128gb to have some future proofing. I figure at this point, it's akin to a mechanic keeping great tools around, when it comes to having this sort of homelab and exposing it for your own uses. And great practice for building the next era of user facing computing that will be around as this proliferates.
I would not buy a 64GB model again, probably, if this were to remain particularly important to me. But I gather memory bandwidth is pretty important here.
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.)
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.
I was interested to see that Qwen3.5-122B-A10B narrowly beat Qwen3.6-27B on Donato Capitella's SWEBench-verified-mini run with a similar 128GB UMA architecture.
Local models are great for a lot of things past just software development. We need to move towards solving other real world problems vs just building software. I've been focused on that with TxtAI (https://github.com/neuml/txtai) for 6 years now.
Something I find really confusing from this post is the MLX versions of the model running much slower. As I understand it, these model versions are meant to take advantage of Apple Silicon and MacOS APIs, and should produce better/faster results. Any insight into what’s happening here?
How does llama.cpp use the GPU efficiently as opposed to MLX?
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.
Llama.cpp uses the GPU very effectively because inference of LLMs is very rudimentary and basically as simple as your GPU memory bandwidth. That's essentially the baseline performance ceiling, with model-specific optimisations like MTP potentially increasing it.
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.
It will run (somewhat slowly) on a five year old M1 Max with 64GB RAM.
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 need to ask, since I have desperately wanted to make Gemma 4 12B work, but im not sure if its the quant (i usually up it to q8, which is a lot higher than iq4_nl that i use for 3.6 27B) or the model itself, but it just starts confusing itself really quickly when I give it coding tasks. And quickly starts failing tool calls.
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.
Strix Halo user here. While Qwen 3.6 27B exhibits remarkable intelligence density, I will still take unsloth's dynamic IQ2_XXS of Minimax M2.7 over Q8_0 Qwen 3.6 27B any day of the week, and this isn't just because of generation speed either. I wrote my own custom harness, and I get hallucinated tool call parameters and bizarre invocations with Q3.6 27B even at Q8_0, but no issues with the IQ2_XXS of M2.7.
Gemma4 31B with MTP enabled is faster and I feel a bit stronger at coding. Either one can run in 32GB VRAM or unified RAM with some tuning (3 bit weights, 8 bit kv cache)
none of these local models are good for development, complete waste of time. nobody has $100k+ hardware sitting around at home to actually run a good model
Spent a week trying to get sensible results out of llama 3. At one point it even simulated doing the work, log output and everything and when I challenged it about the missing artefacts it actually started questioning my intelligence. Seems appropriate for a Zuck enterprise.
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.
I've been running it almost since launch on a 3090 (24gb vram), you really don't need that much. Second hand those are really cheap and i get 50-70 t/s (with MTP at 2), full ctx. IQ4_NL (unsloth) on this model seems suspiciously competent, and after the (by now not so recent) updates to q4 KV on llama.cpp, I just keep going back to it after dsv4pro disappointed me for the 100th time because it gave up on a task.
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.
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...
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.
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)
If you want to run unquantized, you definitely need 128GB.
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.
[1] https://x.com/MiaAI_lab/status/2070859135399182444
[2] https://github.com/MiaAI-Lab/Qwen3.6-27B-NVFP4-vLLM
So like... $2000+ just for the used GPUs? Plus I assume it's considerably more effort to get it working.
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.
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 think you might be a little to into the stew here.
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 eyars. 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 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.
... 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.
Hahhahahahhahahahhahahhahaha
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.
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"
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
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.
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.
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.
Jackrong has a few different ones available depending on what you're trying to do: https://huggingface.co/Jackrong
(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.)
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.
Sent from my 8gb M2 Mac mini.
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.
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.)
>
> --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
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.
https://pi-local-coding-bench.dev
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.
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.
tweaking sampler might help
I'm running the NVFP4 alongside Gemma4 at the same quant on an OEM Spark
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