Claude Code buying advise

[hehe299792458]

Im starting to use Claude Code to create some apps for myself. Im pretty new to vibe coding, so many breaking a lot of things. In particular, I want to separate it from my person system so in case it messes up, or crashes my OS, it doesn't impact my personal computer.

Right now, I have a MacBook Pro M2 Max with 96GB of RAM. What's the 2nd computer dedicated to vibe coding Mac apps I should buy? I'm considering between a M4 Pro Mac mini or even a Mac Studio (downside, it's going to be replaced soon). Or a 13 inch MacBook Air (very portable, but it's a notebook, not sure how I feel about running it 24/7 like a desktop, also limited to 32GB ram / M5 (no pro chip). Or an M5 MacBook Pro (14 inch). Upside: M5 chip like the air, even Pro variants. Downside: it's heavy and duplicative of my existing MacBook Pro

Any suggestions on what I should get? Im trying to write an app that uses the Gemma 4 LLM locally (using Claude Code to write it). Currently trying the E4B version, but may try others like Qwen too. So running the app does take a bit of memory. Curious which computer I should get for this?

 

[Bigwaff]

If you can afford a high memory configuration, why not purchase a Anthropic/OpenAI/etc monthly subscription instead? Certainly would be cheaper in the long run. Also a “clean separation from your person system”.

 

[sevoneone]

If I understand hehe299792458, you are looking to code with Claude (subscription) but what you are building will utilize a local LLM like Gemma 4. You are not looking to run a coding capable model locally. Disregard me if I'm wrong.

If that is the case though... Since you have a M2 Max with 96GB of RAM, you have plenty to run a macOS VM for a sandboxed Claude code development environment, while running Gemma 4 E4B in LM Studio or similar. Your dev environment could then access your local Gemma model via LM Studio's API. Claude gets installed and runs only inside the VM so you don't have to worry about it 'breaking' anything.

Gemma 4 E4B needs a max of around 18GB if you're using the full precision model and the full 128k token context. That means you could easily run it and a macOS VM with 64GB of RAM to build your app in and still have a 14GB buffer for your primary os/apps while you work. That 64 GB is pretty overkill for the VM too, so you could easily trade ram from it toward a different/larger LLM model or give it back to the primary system.

None of the macs you mention, save the the M5 Pro MacBook Pro or Mac Studio are even available with 64GB of RAM currently, and even they would be more limiting. The Studio is backordered until October. An M5 Pro MBP will have around half the GPU cores and 25% less memory bandwidth as your M2 Max, so the LLM performance between the two is likely to be somewhat similar. I wouldn't even consider the base M5 for LLM vs. the M2 Max, the memory bandwidth alone makes it no contest.

It's just my opinion but, of all the options you gave and what you're describing, you already have the most capable Mac for the job, you just need to think through the config for it.