Hi, I'm Chris 👋

There’s a ton of chatter lately around the expense of AI, with headings like “AI’s real cost problem: Using the tech is more expensive than paying human employees”. Wow, maybe our jobs aren’t in danger! Anyway, I’m not here to express an opinion about clankers replacing humans, that’s for another day. Here I’m just wanting to ask and answer the question: Can you simply run AI locally? If so, what’s it like? And just how capable is the 2026 MacBook Neo, really?

So, to start with, here’s a video of one of the newer highly optimised AI models running on my Neo:

As you can see, it’s certainly quick enough to be useful! And arguably more reliable than using AI online where, if their servers are overloaded, you often don’t get an answer at all (I’m looking at you, Gemini…).

How did I do this? Simply by following this recipe:

• LM Studio (lmstudio.ai)
• LM Studio > My Models > Search > E2B > Google > Download options > MLX 4-bit > Download
• Click the ‘downloading’ button to open the downloads, pause and resume when it gets stuck

I tried various other approaches using command-line python, Apple’s MLX library, etc, but none of that worked. LM Studio is good.

SkiFree clone output

So I asked it to generate a SkiFree game. Here’s what it came up with:

As you can see, nowhere near SkiFree. However, in its defence:

• It did generate working HTML + JS just fine.
• It is some sort of game, to be fair.
• I didn’t re-prompt it several times, which you would if you were actually using it.
• It “admitted” that it was just making a jumping game, not a SkiFree clone.
• It output 21 tokens/sec.

Plasma output

I then asked it to generate a ‘plasma’ graphics effect. Here’s what it made:

Again, not quite a traditional plasma effect, but still something fascinating!

Conclusion

So, in conclusion, even if you have a cheap-as-chips Neo, you can run AI locally just fine, and it’s obvious not as good as the frontier commercial models, but it’s better than you’d think, and plenty fine for the way I use AI: like a supercharged search.

Give it a try! May i recommend LM Studio and the <= 4-bit MLX models.

Thanks for reading, I pinky promise this was written by a human, not AI, hope you found this helpful, at least a tiny bit, God bless!

Photo from unsplash.com/@allthestories