Quick question: Should I stick with my M4 Max or grab a Corsair AI Workstation 300 for local LLM stuff?
Posted by SnooCrickets7501@reddit | LocalLLaMA | View on Reddit | 46 comments
So I already have a Mac Studio M4 Max (return window still available)with 64GB RAM, but I’m eyeing the Corsair AI Workstation 300 (Ryzen AI Max+ 395, 96 VRAM out of 128GB, $3,250). Both seem decent for running models locally with Ollama.
The Corsair has twice the RAM, which feels like it’d be better for bigger models, but I’m not sure if that justifies another 450 + having to switch from Mac to Windows.Other stuffs like speed and stability. also say mac studio is now allowing Nvidia and AMD gpu upgrades would u guys strongly recommend i switch or this ain’t thay good
Im doing some AI projects and experimenting with local tools. Has anyone used either of these for local LLM inference? Is the M4 Max 64GB enough or does it start struggling with larger models?
Would love to hear from people who’ve actually worked with these machines.
SnooCrickets7501@reddit (OP)
Initiated the return guys
dilberx@reddit
Buying corsair?
SnooCrickets7501@reddit (OP)
no buying nimo with the exact same specs but 700$ less because its a smaller brand saw some good reviews and it has 90 days no questions asked policy!
dilberx@reddit
Why not mac? I see its best value. Best performance, best energy efficiency.
SnooCrickets7501@reddit (OP)
i have been using mac for the past 2 weeks over 100 hours it not best performance for what i am doing and mac os is a shitty environment for developers i can’t customise a lot of things a Lot which i can do in linux mac os takes up 11 gb of the vram of 64 while linux only takes 4-5gb i am getting twice the ram for 400$ less while getting out of mac os yes mac seems more premium but it just seems premium cause of the flashy ui and it does have faster token generation and better cooling thats it mac is not for actual developers i am studying ML/AI in college but yea mac might be better for other cause not for me. Also think the whole mac mini trend is just created by non developers business owners not actual developers who wants to learn this hope this helps but mac might definitely be better for your use case depending on how u use it.
ducksoup_18@reddit
This thing? https://www.nimopc.com/products/nimo-ai-mini-pc-amd-ryzen-ai-max-395-128gb-ram Hadnt heard of it. Might need to look into it. That price tag is a bit more palpable.
SnooCrickets7501@reddit (OP)
your link don’t work on me but for me it cost 2450 total with taxes and shipping
SnooCrickets7501@reddit (OP)
yesss
CATLLM@reddit
The mac studios are terrible for llm and everyone benches them with 0 context filled and think its fast.
Creepy-Bell-4527@reddit
Compared to a Strix Halo, which is an equally bad iGPU with significantly slower memory, they're fast.
CATLLM@reddit
Strix halo still performant in thr 32k - 64k context range which is where you’ll fill it to for regular work. Mac m3 m4 all tank after 32k context.
Creepy-Bell-4527@reddit
Unless you're frontloading all 32k tokens, prefix caching almost completely eliminates the slowdown.
CATLLM@reddit
Bro, for real money paying work, agentic workflows 32k fills up easily. I work with 128k for qwen3.5 122b.
If you are using it for you AI waifu then you’re probably fine with 8k.
Creepy-Bell-4527@reddit
Front loading and progressively filling aren’t the same thing. Prefix caching makes the latter fast, and the latter is what agentic coding does.
I also handle 122b at 100+k context without breaking a sweat.
CATLLM@reddit
It ain’t fast compared to Strix halo and dgx spark. I don’t what your definition of “fast” is.
Creepy-Bell-4527@reddit
Don’t pair the DGX Spark and Strix Halo like they’re peers in terms of performance. They aren’t.
The M3 Ultra 60c running 122b @ Q4 gets 21 tok/s at 64k. The Halo gets 15tok/s at 50k. The Spark leaves them both in the dust.
Stop chatting out your ass.
CATLLM@reddit
Bro a 4bit mlx is not the same as a q4 dynamic gguf in both speed and quality. A 4bit mlx is more like a q3xs gguf.
This is what im saying, Mac Studio bros running LLMs are not making real money with AI so they don’t know what they don’t know.
Creepy-Bell-4527@reddit
😂 cry harder.
CATLLM@reddit
LMAO 15t/s . Exactly proves my point anyone using Mac Studios for LLM are not making money with AI cause that shit is not usable at 15t/s
Creepy-Bell-4527@reddit
Still .1 t/s faster than the 395+, at 2 bit higher quant and 14k higher context 😘 And over twice the PP speed.
CATLLM@reddit
Lmao i don’t have a 395+ cause i don’t want to deal with that stupid shit. I have M5 max, 2x dgx spark and rtx 6000 pro.
JacketHistorical2321@reddit
Idiot
CATLLM@reddit
Max cope
SnooCrickets7501@reddit (OP)
ohh okay so corsair wins in this case?
CATLLM@reddit
100% the corsair wins.
KFSys@reddit
If you’re mostly experimenting, I’d probably just stick with what you have for now. The M4 Max with 64GB is already pretty capable, and you can run a lot of stuff locally before it becomes a real limitation.
The Corsair setup looks nice on paper, especially with more RAM/VRAM, but it’s a pretty big jump in cost and you’re also switching ecosystems. Not sure it’s worth it unless you already know you’re hitting limits with your current setup.
What worked better for me was just using cloud GPUs when I needed more power instead of overcommitting to hardware. I’ve spun up things like DigitalOcean GPU instances for heavier runs and then just shut them down after. Way cheaper if you’re not using it all the time.
SnooCrickets7501@reddit (OP)
i wound not say im experimenting anymore i was now im building Multi LLM agent workflows that cab automate any work for me i need and help run a business i got one running in mac but i always get border-lined by memory because i have to use multiple apps for automation like n8n appolo serp obsidian
gabrielesilinic@reddit
If you want to spend and do stuff for real build a PC, at least you will be able to upgrade it as time goes on. The other thing you are buying is basically a mac again but x86 and marginally better software support sometimes.
But like a mac you will have to trash it in a few years or if you outgrow it. Instead if you plan for a PC properly even buying an AMD GPU may serve you better over time. Though I admit that GPUs with enough vram are not cheap at all.
SnooCrickets7501@reddit (OP)
true im a huge pc guy i got and 3070ti but those are for gaming with my workload i wont be able to game there and i can upgrade the corsair witha igpu deck upto 2 gpus
HopePupal@reddit
that's a Strix Halo box. they're really common here because they're cheap and have up to 128 GB unified memory (you can use more than 96 GB of it, just not on Windows). i have one. they're not the fastest but they can run MiniMax.
SnooCrickets7501@reddit (OP)
whats ur suggestion then corsair is better or any other windows/linux based machine?
Automatic-Arm8153@reddit
Bosgame m5 is usually the cheapest 128gb strip halo, check that one out.
Both Mac and strix halo devices will be relatively slow for AI stuff though. However they are good beginner devices to get you into the game.
When you eventually begin caring about speed start looking at nvidia GPU options. For now I’d get a strict halo, and drop windows. Use Linux.
It’s not that hard anymore since you can ask AI for help using it. I personally run Debian across my machines.
SnooCrickets7501@reddit (OP)
bossgame m5 would be better than corsair AI 300 even though its cheaper thats pretty good deal
SnooCrickets7501@reddit (OP)
got nimo thats cheaper and same exact specs got nimo for 2500 total with shipping
Automatic-Arm8153@reddit
Wait are you saying the Corsair is cheaper for you?
If it’s cheaper then it’s better. Since they run the same internals both 128gb. Get whichever device is cheaper.
SnooCrickets7501@reddit (OP)
no corsair is same for 1tb but its not available 4 tb is like 450 extra
Royale_AJS@reddit
1TB should be plenty for models unless you’re constantly hopping around.
SnooCrickets7501@reddit (OP)
i know thats why i want to spend less but only 4tb is available
Iiiwabibitoiii@reddit
If you get the strix halo, use fedora and the strix halo toolboxes. It will save you a lot of headaches.
inthesearchof@reddit
Return the M4 Max and hold out for the Mac Studio M5. short term pain for long term satisfaction.
SnooCrickets7501@reddit (OP)
apple releasing m5 on mac mini and studio is very thin they usually skip generation chips
JacketHistorical2321@reddit
Ya, and where is the M4 ultra??
Ok_Technology_5962@reddit
I would go with mac. The developer environment on ai just exploded. Features coming like no tomorrow. Also be mindful in case that system restricts vram allocation in that 128 gigs
Iiiwabibitoiii@reddit
I have a strix halo and the Mac Studio m4 max. The Mac has much better memory bandwidth.
Creepy-Bell-4527@reddit
The 8060S (Max+ 395 GPU) and the M4 Max GPU are very similar in performance (similar prompt processing speeds), but the unified memory is significantly faster (meaning faster token generation speeds), and Mac still has better tooling than the 395.
I would however suggest picking a Strix Halo box with Oculink. Mac eGPU support is still terrible, Oculink isn't.
Powerful_Evening5495@reddit
Corsair AI Workstation 300 1000%