If this is your first time using open weight models right after release, know that there are always bugs in the early implementations and even quantizations.
Every project races to have support on launch day so they don’t lose users, but the output you get may not be correct. There are already several problems being discovered in tokenizer implementations and quantizations may have problems too if they use imatrix.
So you’re going to see a lot of “I tried it but it sucks because it can’t even do tool calls” and other reports about how the models don’t work at all in the coming weeks from people who don’t realize they were using broken implementations.
If you want to try cutting edge open models you need to be ready to constantly update your inference engine and check your quantization for updates and re-download when it’s changed. The mad rush to support it on launch day means everything gets shipped as soon as it looks like it can produce output tokens, not when it’s tested to be correct.
You seem like you know what you're talking about... what inference engine should I use? (linux, 4090)
I keep having "I tried it but it sucks" issues mostly around tool calling and it's not clear if it's the model or ollama. And not one model in particular, any of them really.
I've had really good success with LMStudio and GLM 4.7 Flash and the Zed editor which has a baked in integration with LMStudio. I am able to one-shot whole projects this way, and it seems to be constantly improving. Some update recently even allowed the agent to ask me if it can do a "research" phase - so it'll actually reach out to website and read docs and code from github if you allow it. GLM 4.7 flash has been the most adept at tool calling I've found, but the Qwen 3 and 3.5 models are also fairly good, though run into more snags than I've seen with GLM 4.7 flash.
For the specific issue parent is talking about, you really need to give various tools a try yourself, and if you're getting really shit results, assume it's the implementation that is wrong, and either find an existing bug tracker issue or create a new one.
Same thing happened when GPT-OSS launched, bunch of projects had "day-1" support, but in reality it just meant you could load the model basically, a bunch of them had broken tool calling, some chat prompt templates were broken and so on. Even llama.cpp which usually has the most recent support (in my experience) had this issue, and it wasn't until a week or two after llama.cpp that GPT-OSS could be fairly evaluated with it. Then Ollama/LM Studio updates their llama.cpp some days after that.
So it's a process thing, not "this software is better than that", and it heavily depends on the model.
I don’t know if any of engines are fully tested yet.
For new LLMs I get in the habit of building llama.cpp from upstream head and checking for updated quantizations right before I start using it. You can also download llama.cpp CI builds from their release page but on Linux it’s easy to set up a local build.
If you don’t want to be a guinea pig for untested work then the safe option would be to wait 2-3 weeks
just use openrouter or google ai playground for the first week till bugs are ironed out. You still learn the nuances of the model and then yuu can switch to local. In addition you might pickup enough nuance to see if quantization is having any effect
Huge Claude user here… can someone help me set some realistic expectations if I bought a Mac mini and spun one up? I use Claude primarily for dev work and Home Lab projects. Are the open models good enough to run locally and replace the Claude workload? Or am I better off with my $20/mo Claude subscription?
They are good for small tasks but you would not be able to use it like you use Claude and most likely be disappointed. But also, I do not know how you use claude.
There are many services online which offer hosted services for these models, my advice for anyone who is thinking about buying hardware to self host this is to try those first, that way you can get an impression of the capabilities and limitations of those models before you commit to buying hardware
M5 air here with 32gb ram and 10/10 cores. Anyone got some luck with mlx builds on oMLX so far? Not at my machine right now and would love to know if these models already work including tool calling
The latest release v0.3.2 has partial support, generation is supported but not all special tokens are handled. I've done some personal testing to add tool calling and <|channel> thinking support. https://github.com/Yukon/omlx
I tested briefly with a MacBook Pro m4 with 36gb. Run in LM Studio with open code as the frontend and it failed over and over on tool calls. Switched back to qwen. Anyone else on similar setup have better luck?
I failed to run in LM Studio on M5 with 32gb at even half max context. Literally locked up computer and had to reboot.
Ran gemma-4-26B-A4B-it-GGUF:Q4_K_M just fine with llama.cpp though. First time in a long time that I have been impressed by a local model. Both speed (~38t/s) and quality are very nice.
Even with the latest version of LM Studio and the latest runtimes I find that tool use fails 100% of the time with the following error: Error rendering prompt with jinja template: "Cannot apply filter "upper" to type: UndefinedValue".
EDIT: The issue is addressed in LM Studio 0.4.9 (build 1), which auto-update wasn't picking up for me for some reason.
Yes, you can use it for local coding. Most harnesses can be pointed at a local endpoint which provides an OpenAI compatible API, though I've had some trouble using recent versions of Codex with llama.cpp due to an API incompatibility (Codex uses the newer "responses" API, but in a way that llama.cpp hasn't fully supported).
I personally prefer Pi as I like the fact that it's minimalist and extensible. But some people just use Claude Code, some OpenCode, there are a ton of options out there and most of them can be used with local models.
It needs to support tool calling and many of the quantized ggufs don't so you have to check.
I've got a workaround for that called petsitter where it sits as a proxy between the harness and inference engine and emulates additional capabilities through clever prompt engineering and various algorithms.
They're abstractly called "tricks" and you can stack them as you please.
Why is ollama so many people’s go-to? Genuinely curious, I’ve tried it but it feels overly stripped down / dumbed down vs nearly everything else I’ve used.
Lately I’ve been playing with Unsloth Studio and think that’s probably a much better “give it to a beginner” default.
Ollama is good enough to dabble with, and getting a model is as easy as ollama pull <model name> vs figuring it out by yourself on hugging face and trying to make sense on all the goofy letters and numbers between the forty different names of models, and not needing a hugging face account to download.
So you start there and eventually you want to get off the happy path, then you need to learn more about the server and it's all so much more complicated than just using ollama. You just want to try models, not learn the intricacies of hosting LLMs.
to be fair, llama.cpp has gotten much easier to use lately with llama-server -hf <model name>. That said, the need to compile it yourself is still a pretty big barrier for most people.
Ollama got some first-mover advantage at the time when actually building and git pulling llama.cpp was a bit of a moat. The devs' docker past probably made them overestimate how much they could lay claim to mindshare. However, no one really could have known how quickly things would evolve... Now I mostly recommend LM-studio to people.
LM Studio has been around longer. I’ve used it since three years ago. I’d also agree it is generally a better beginner choice then and now.
Unsloth Studio is more featureful (well integrated tool calling, web search, and code execution being headline features), and comes from the people consistently making some of the best GGUF quants of all popular models. It also is well documented, easy to setup, and also has good fine-tuning support.
Ollama user with the opposite question -- why not? What am I missing out on? I'm using it as the backend for playing with other frontend stuff and it seems to work just fine.
And as someone running at 16gb card, I'm especially curious as to if I'm missing out on better performance?
Ollama has had bad defaults forever (stuck on a default CTX of 2048 for like 2 years) and they typically are late to support the latest models vs llamacpp. Absolutely no reason to use it in 2026.
Ollama's org had people flood various LLM/programming related Reddits and Discords and elsewhere, claiming it was an 'easy frontend for llama.cpp', and tricked people.
Only way to win is to uninstall it and switch to llama.cpp.
There is virtually no reason to use Ollama over LM Studio or the myriad of other alternatives.
Ollama is slower and they started out as a shameless llama.cpp ripoff without giving credit and now they “ported” it to Go which means they’re just vibe code translating llama.cpp, bugs included.
Do y'all mean backend or the Ollama frontend or both? I find it trivially easy to sub in my local Ollama api thing in virtually all of the interesting frontend things. I'm quite curious about the "why not Ollama" here.
I really like LM Studio when I can use it under Windows but for people like me with Intel Macs + AMD gpu ollama is the only option because it can leverage the gpu using MoltenVK aka Vulkan, unofficially. We're still testing it, hoping to get the Vulkan support in the main branch soon. It works perfectly for single GPUs but some edge cases when using multiple GPUs are unsupported until upstream support from MoltenVK comes through. But yeah, I agree, it wasn't cool to repackage Georgi's work like that.
I don't think it does, but llama.cpp does, and can load models off HuggingFace directly (so, not limited to ollama's unofficial model mirror like ollama is).
Yes, they introduced that Golang rewrite precisely to support the visual pipeline and other things that weren't in llama.cpp at the time. But then llama.cpp usually catches up and Ollama is just left stranded with something that's not fully competitive. Right now it seems to have messed up mmap support which stops it from properly streaming model weights from storage when doing inference on CPU with limited RAM, even as faster PCIe 5.0 SSDs are finally making this more practical.
The project is just a bit underwhelming overall, it would be way better if they just focused on polishing good UX and fine-tuning, starting from a reasonably up-to-date version of what llama.cpp provides already.
In some places in the source code they claim sole ownership of the code, when it is highly derivative of that in llama.cpp (having started its life as a llama.cpp frontend). They keep it the same license, however, MIT.
There is no reason to use Ollama as an alternative to llama.cpp, just use the real thing instead.
If it’s MIT code derived from MIT code, in what way is its openness ”quasi”? Issues of attribution and crediting diminish the karma of the derived project, but I don’t see how it diminishes the level of openness.
I've benchmarked this on an actual Mac Mini M4 with 24 GB of RAM, and averaged 24.4 t/s on Ollama and 19.45 t/s on LM Studio for the same ~10 GB model (gemma4:e4b), a difference which was repeated across three runs and with both models warmed up beforehand. Unless there is an error in my methodology, which is easy to repeat[1], it means Ollama is a full 25% faster. That's an enormous difference. Try it for yourself before making such claims.
[1] script at: https://pastebin.com/EwcRqLUm but it warms up both and keeps them in memory, so you'll want to close almost all other applications first. Install both ollama and LM Studio and download the models, change the path to where you installed the model. Interestingly I had to go through 3 different AI's to write this script: ChatGPT (on which I'm a Pro subscriber) thought about doing so then returned nothing (shenanigans since I was benchmarking a competitor?), I had run out of my weekly session limit on Pro Max 20x credits on Claude (wonder why I need a local coding agent!) and then Google rose to the challenge and wrote the benchmark for me. I didn't try writing a benchmark like this locally, I'll try that next and report back.
It depends on the hardware, backend and options. I've recently tried running some local AIs (Qwen3.5 9B for the numbers here) on an older AMD 8GB VRAM GPU (so vulkan) and found that:
llama.cpp is about 10% faster than LM studio with the same options.
LM studio is 3x faster than ollama with the same options (~13t/s vs ~38t/s), but messes up tool calls.
Ollama ended up slowest on the 9B, Queen3.5 35B and some random other 8B model.
Note that this isn't some rigorous study or performance benchmarking. I just found ollama unnaceptably slow and wanted to try out the other options.
In case someone would like to know what these are like on this hardware, I tested Gemma 4 32b (the ~20 GB model, the largest Gemma model Google published) and Gemma 4 gemma4:e4b (the ~10 GB model) on this exact setup (Mac Mini M4 with 24 GB of RAM using Ollama), I livestreamed it:
The ~10 GB model is super speedy, loading in a few seconds and giving responses almost instantly. If you just want to see its performance, it says hello around the 2 minute mark in the video (and fast!) and the ~20 GB model says hello around 5 minutes 45 seconds in the video. You can see the difference in their loading times and speed, which is a substantial difference. I also had each of them complete a difficult coding task, they both got it correct but the 20 GB model was much slower. It's a bit too slow to use on this setup day to day, plus it would take almost all the memory. The 10 GB model could fit comfortably on a Mac Mini 24 GB with plenty of RAM left for everything else, and it seems like you can use it for small-size useful coding tasks.
Isn't 26 tok/s quite usable for a claw-like agent though? You can chat with it on a IM platform and get notified as soon as it replies, you're not dependent on real-time quick interaction.
By desk you mean that "Mac mini"? Because it is pricey. In my country it is 1000 USD (from Apple for basic M4 with 24GB). My desk was 1/5th of that price.
And considering that this Mac mini won't be doing anything else is there a reason why not just buy subscription from Claude, OpenAI, Google, etc.?
Are those open models more performant compared to Sonnet 4.5/4.6? Or have at least bigger context?
Right now, open models that run on hardware that costs under $5000 can get up to around the performance of Sonnet 3.7. Maybe a bit better on certain tasks if you fine tune them for that specific task or distill some reasoning ability from Opus, but if you look at a broad range of benchmarks, that's about where they land in performance.
You can get open models that are competitive with Sonnet 4.6 on benchmarks (though some people say that they focus a bit too heavily on benchmarks, so maybe slightly weaker on real-world tasks than the benchmarks indicate), but you need >500 GiB of VRAM to run even pretty aggressive quantizations (4 bits or less), and to run them at any reasonable speed they need to be on multi-GPU setups rather than the now discontinued Mac Studio 512 GiB.
The big advantage is that you have full control, and you're not paying a $200/month subscription and still being throttled on tokens, you are guaranteed that your data is not being used to train models, and you're not financially supporting an industry that many people find questionable. Also, if you want to, you can use "abliterated" versions which strip away the censoring that labs do to cause their models to refuse to answer certain questions, or you can use fine-tunes that adapt it for various other purposes, like improving certain coding abilities, making it better for roleplay, etc.
You don't need that much VRAM to run the very largest models, these are MoE models where only a small fraction is being computed with at any given time. If you plan to run with multiple GPUs and have enough PCIe lanes (such as with a proper HEDT platform) CPU-GPU transfers start to become a bit less painful. More importantly, streaming weights from disk becomes feasible, which lets you save on expensive RAM. The big labs only avoid this because it costs power at scale compared to keeping weights in DRAM, but that aside it's quite sound.
I have the same setup (M4 Pro, 24GB). The e4b model is surprisingly snappy for quick tasks. The full 26B is usable but not great — loading time alone is enough to break your flow.
Re: subscriptions vs local — I use both. Cloud for the heavy stuff, local for when I'm iterating fast and don't want to deal with rate limits or network hiccups.
The article has a few good tips for using Ollama. Perhaps it should note that the Gemma 4 models are not really trained for strong performance with coding agents like OpenCode, Claude Code, pi, etc. The Gemma 4 models are excellent for applications requiring tool use, data extraction to JSON, etc. I asked Gemini Pro about this earlier and Gemini Pro recommended qwen 3.5 models specifically for coding, and backed that up with interesting material on training. This makes sense, and is something that I do: use strong models to build effective applications using small efficient models.
> I asked Gemini Pro about this earlier and Gemini Pro recommended qwen 3.5 models specifically for coding, and backed that up with interesting material on training.
The Gemma models were literally released yesterday. You can’t ask LLMs for advice on these topics and get accurate information.
Please don’t repeat LLM-sourced answers as canonical information
It's not just LLM sourced though, folks have literally tried this after the release with the 26A4B model and it wasn't very good. Maybe the dense ~31B model is worthwhile though.
I agree with your criticism. I should have simply said that I had good results with gemma 4 tool use, and agentic coding with gemma 4 didn’t yet work well for me.
I spent two hours doing my own research before asking for Gemini’s analysis, which reinforced my own opinion that the gemini models historically have not been trained and target for agentic coding use.
Have you tried using the new Gemma 4 models with agentic coding tools?If you do, you might end up agreeing with me.
I wasn’t very clear, sorry. By my ‘own research’ I meant spending 90 minutes experimenting with Gemma 4 models for tool use (good results!) and a half hour using with pi and OpenCode (I didn’t get good results, yet.)
Oh yeah absolute genius. I asked GPT-2 about Claude Opus 4.6 and it said “this is not a recommendation. You might get some benefits from Opus… but this is not what you want”. Damn, real wisdom from the OG there. What a legend
Every project races to have support on launch day so they don’t lose users, but the output you get may not be correct. There are already several problems being discovered in tokenizer implementations and quantizations may have problems too if they use imatrix.
So you’re going to see a lot of “I tried it but it sucks because it can’t even do tool calls” and other reports about how the models don’t work at all in the coming weeks from people who don’t realize they were using broken implementations.
If you want to try cutting edge open models you need to be ready to constantly update your inference engine and check your quantization for updates and re-download when it’s changed. The mad rush to support it on launch day means everything gets shipped as soon as it looks like it can produce output tokens, not when it’s tested to be correct.
I keep having "I tried it but it sucks" issues mostly around tool calling and it's not clear if it's the model or ollama. And not one model in particular, any of them really.
Same thing happened when GPT-OSS launched, bunch of projects had "day-1" support, but in reality it just meant you could load the model basically, a bunch of them had broken tool calling, some chat prompt templates were broken and so on. Even llama.cpp which usually has the most recent support (in my experience) had this issue, and it wasn't until a week or two after llama.cpp that GPT-OSS could be fairly evaluated with it. Then Ollama/LM Studio updates their llama.cpp some days after that.
So it's a process thing, not "this software is better than that", and it heavily depends on the model.
For new LLMs I get in the habit of building llama.cpp from upstream head and checking for updated quantizations right before I start using it. You can also download llama.cpp CI builds from their release page but on Linux it’s easy to set up a local build.
If you don’t want to be a guinea pig for untested work then the safe option would be to wait 2-3 weeks
There are many services online which offer hosted services for these models, my advice for anyone who is thinking about buying hardware to self host this is to try those first, that way you can get an impression of the capabilities and limitations of those models before you commit to buying hardware
1: https://github.com/bolyki01/localllm-gemma4-mlx
This is how all open weight model launches go.
Ran gemma-4-26B-A4B-it-GGUF:Q4_K_M just fine with llama.cpp though. First time in a long time that I have been impressed by a local model. Both speed (~38t/s) and quality are very nice.
EDIT: The issue is addressed in LM Studio 0.4.9 (build 1), which auto-update wasn't picking up for me for some reason.
https://github.com/ggml-org/llama.cpp/issues/21347#issuecomm...
I personally prefer Pi as I like the fact that it's minimalist and extensible. But some people just use Claude Code, some OpenCode, there are a ton of options out there and most of them can be used with local models.
I've got a workaround for that called petsitter where it sits as a proxy between the harness and inference engine and emulates additional capabilities through clever prompt engineering and various algorithms.
They're abstractly called "tricks" and you can stack them as you please.
https://github.com/day50-dev/Petsitter
You can run the quantized model on ollama, put petsitter in front of it, put the agent harness in front of that and you're good to go
If you have trouble, file bugs. Please!
Thank you
Lately I’ve been playing with Unsloth Studio and think that’s probably a much better “give it to a beginner” default.
So you start there and eventually you want to get off the happy path, then you need to learn more about the server and it's all so much more complicated than just using ollama. You just want to try models, not learn the intricacies of hosting LLMs.
What does unsloth-studio bring on top?
Unsloth Studio is more featureful (well integrated tool calling, web search, and code execution being headline features), and comes from the people consistently making some of the best GGUF quants of all popular models. It also is well documented, easy to setup, and also has good fine-tuning support.
And as someone running at 16gb card, I'm especially curious as to if I'm missing out on better performance?
Ollama's org had people flood various LLM/programming related Reddits and Discords and elsewhere, claiming it was an 'easy frontend for llama.cpp', and tricked people.
Only way to win is to uninstall it and switch to llama.cpp.
Ollama is slower and they started out as a shameless llama.cpp ripoff without giving credit and now they “ported” it to Go which means they’re just vibe code translating llama.cpp, bugs included.
There is no reason to ever use ollama.
I just checked their docs and can't see anything like it.
Did you mistake the command to just download and load the model?
Actually that shouldn't be a question, you clearly did.
Hint: it also opens Claude code configured to use that model
And didn't Ollama independently ship a vision pipeline for some multimodal models months before llama.cpp supported it?
The project is just a bit underwhelming overall, it would be way better if they just focused on polishing good UX and fine-tuning, starting from a reasonably up-to-date version of what llama.cpp provides already.
Hmm, the fact that Ollama is open-source, can run in Docker, etc.?
In some places in the source code they claim sole ownership of the code, when it is highly derivative of that in llama.cpp (having started its life as a llama.cpp frontend). They keep it the same license, however, MIT.
There is no reason to use Ollama as an alternative to llama.cpp, just use the real thing instead.
I've benchmarked this on an actual Mac Mini M4 with 24 GB of RAM, and averaged 24.4 t/s on Ollama and 19.45 t/s on LM Studio for the same ~10 GB model (gemma4:e4b), a difference which was repeated across three runs and with both models warmed up beforehand. Unless there is an error in my methodology, which is easy to repeat[1], it means Ollama is a full 25% faster. That's an enormous difference. Try it for yourself before making such claims.
[1] script at: https://pastebin.com/EwcRqLUm but it warms up both and keeps them in memory, so you'll want to close almost all other applications first. Install both ollama and LM Studio and download the models, change the path to where you installed the model. Interestingly I had to go through 3 different AI's to write this script: ChatGPT (on which I'm a Pro subscriber) thought about doing so then returned nothing (shenanigans since I was benchmarking a competitor?), I had run out of my weekly session limit on Pro Max 20x credits on Claude (wonder why I need a local coding agent!) and then Google rose to the challenge and wrote the benchmark for me. I didn't try writing a benchmark like this locally, I'll try that next and report back.
llama.cpp is about 10% faster than LM studio with the same options.
LM studio is 3x faster than ollama with the same options (~13t/s vs ~38t/s), but messes up tool calls.
Ollama ended up slowest on the 9B, Queen3.5 35B and some random other 8B model.
Note that this isn't some rigorous study or performance benchmarking. I just found ollama unnaceptably slow and wanted to try out the other options.
https://www.youtube.com/live/G5OVcKO70ns
The ~10 GB model is super speedy, loading in a few seconds and giving responses almost instantly. If you just want to see its performance, it says hello around the 2 minute mark in the video (and fast!) and the ~20 GB model says hello around 5 minutes 45 seconds in the video. You can see the difference in their loading times and speed, which is a substantial difference. I also had each of them complete a difficult coding task, they both got it correct but the 20 GB model was much slower. It's a bit too slow to use on this setup day to day, plus it would take almost all the memory. The 10 GB model could fit comfortably on a Mac Mini 24 GB with plenty of RAM left for everything else, and it seems like you can use it for small-size useful coding tasks.
brew install llama.cpp
use the inbuilt CLI, Server or Chat interface. + Hook it up to any other app
And considering that this Mac mini won't be doing anything else is there a reason why not just buy subscription from Claude, OpenAI, Google, etc.?
Are those open models more performant compared to Sonnet 4.5/4.6? Or have at least bigger context?
You can get open models that are competitive with Sonnet 4.6 on benchmarks (though some people say that they focus a bit too heavily on benchmarks, so maybe slightly weaker on real-world tasks than the benchmarks indicate), but you need >500 GiB of VRAM to run even pretty aggressive quantizations (4 bits or less), and to run them at any reasonable speed they need to be on multi-GPU setups rather than the now discontinued Mac Studio 512 GiB.
The big advantage is that you have full control, and you're not paying a $200/month subscription and still being throttled on tokens, you are guaranteed that your data is not being used to train models, and you're not financially supporting an industry that many people find questionable. Also, if you want to, you can use "abliterated" versions which strip away the censoring that labs do to cause their models to refuse to answer certain questions, or you can use fine-tunes that adapt it for various other purposes, like improving certain coding abilities, making it better for roleplay, etc.
Re: subscriptions vs local — I use both. Cloud for the heavy stuff, local for when I'm iterating fast and don't want to deal with rate limits or network hiccups.
The Gemma models were literally released yesterday. You can’t ask LLMs for advice on these topics and get accurate information.
Please don’t repeat LLM-sourced answers as canonical information
Everyone hated Qwen3.5 at launch too because so many implementations were broken and couldn’t do tool calling.
You need to ignore social media “I tried this and it sucks” echo chambers for new model releases.
Have you tried using the new Gemma 4 models with agentic coding tools?If you do, you might end up agreeing with me.