Local AI for coding and automation: why spend thousands on models weaker than the cloud ones?

For coding and automation, local AI pays off little today. The most powerful consumer hardware handles models up to 200 billion parameters. The best open weight model, Kimi K3, has 2.8 trillion. Spending thousands of euros to run a model at home that is less capable than the ones I use in the cloud means paying more to get a worse result.

More and more professionals are investing serious money in machines for local AI. I struggle to see the advantage, at least for the work I do every day: programming and building automations.

How big a model can local hardware actually run?

The August 2026 announcements give a clear picture of the limits.

• Apple Mac Studio with M5 Ultra: announced on August 25, 2026, it goes up to 512GB of unified memory and starts at $5,499. Apple says it can run "enormous LLMs entirely on device". The 512GB configuration only arrives in late October, and its price is not public yet.
• Xiaomi AI Cube: it is a prototype and is not on sale. Its XRING D100 chip supports up to 160GB of unified memory and models up to 200 billion parameters. Commercial use of the chip is planned for 2027.

Two hundred billion parameters sounds like a huge number. The problem shows up when you compare it with the models you actually need for programming.

Why don’t the best open weight models fit on a local machine?

Kimi K3 by Moonshot AI is the open weight model closest to the best closed models such as Fable 5. It has 2.8 trillion parameters: 14 times the 200 billion supported by Xiaomi’s chip.

So Kimi K3 does not run on Xiaomi’s AI Cube. It does not even fit whole in the memory of a single 512GB Mac Studio, a configuration that at the end of August still had no release date.

There are techniques to run Kimi K3 on Apple Silicon anyway, such as weight streaming and expert pruning. I covered them in Kimi K3 running locally on Apple Silicon. They are interesting experiments, with trade-offs in speed or model completeness, and they are still far from a setup for daily work.

Why does a less capable model cost me time?

For me, the advantage of using AI is delegating as much work as possible. A less capable model makes mistakes more often. If it makes mistakes more often, I have to check its work more often.

Every extra check is time I spend reviewing the work of my digital colleague instead of moving forward. At that point the advantage of using AI shrinks, right in the part of the work where it should matter most.

The math becomes this: tens of thousands of euros in hardware to get a model that needs more supervision than the one I already have in the cloud.

Does privacy justify local AI?

Privacy is the obvious answer, and it is a real argument. Some companies have policies that forbid sending data to American cloud services. In those cases a local model is the only way.

The question is how many professionals actually have that requirement. In my view, today, a minority.

There is also a paradox. If I want to use a top-tier open weight model like Kimi K3, I can’t run it in any practical way on a single machine. I still have to use it in the cloud, through a provider. And the data leaves my computer anyway.

What do I use for coding and automation?

For my daily work I use Claude Code and Codex with the top models. I pay for the subscription and use them for everything I need.

I don’t care where the AI runs. I care that it produces output, makes my work easier and lets me get more done in the same time.

Maybe I’m missing something. If you work with local models for programming or automation, your use case is exactly what I want to understand: tell me about it in the comments under the Short.