Your MacBook is secretly a supercomputer. But most AI tools treat it like a glorified calculator. That's the takeaway from a new benchmark showdown between two Apple Silicon-optimized engines: Kapitsa and Splash. On a MacBook Pro M4 Pro with 48GB of RAM, Kapitsa generated code at a blistering 65 tokens per second, while Splash managed only about 50. That's not a marginal win—it's a blowout. And remember, these are free tools running on hardware you already own. No cloud, no subscription, no waiting for a server farm to respond.
But here's the twist: Kapitsa is version 0.2. It's early. It's buggy. You have to install it yourself, which means terminal commands and troubleshooting. Splash, meanwhile, is at version 1.4.0—more mature, more stable. So the classic trade-off applies: bleeding-edge speed versus peace of mind. For developers who live in their terminal, Kapitsa is a no-brainer. For everyone else, it's a taste of what's coming.
The real story isn't the numbers. It's the shift. For years, running serious AI locally meant buying a gaming PC with a massive GPU. Apple Silicon changed that equation. Now, a MacBook Pro M4 Pro can run a 27-billion-parameter model (like Qwen 3.8) at speeds that make cloud APIs feel sluggish. And the M1 Pro with 16GB? It still holds its own. That's like finding out your old sedan can keep up with a sports car on the highway.
Who wins? Apple. Every time a tool like Kapitsa makes local AI faster, it sells more Macs. Who loses? Cloud providers who charge by the token. If you can get 65 tokens per second on your laptop, why pay for a remote server? The catch: local models are still smaller and less capable than the biggest cloud models. But they're catching up fast.
The benchmark also tested multiple model sizes—4B, 27B, and 35B—all quantized to 4-bit. That's the sweet spot for local AI: small enough to fit in memory, big enough to be useful. And both engines support MLX (with MTP), Apple's machine learning framework. This isn't a hobbyist hack; it's a serious ecosystem.
So what should you do? If you're a developer, grab Kapitsa and file bug reports. If you're not, wait for version 1.0. But know this: the gap between local and cloud AI is shrinking. Trends suggest that within a year, the default way to use AI might be on your own machine. The question is: will you be the one waiting for the cloud, or the one already coding at 65 tokens per second?