July 28, 2026
Mac attack, snail-speed edition
Running Kimi K3 on a M1 Mac
This giant AI ran on a Mac, and the comments instantly asked: but... why so slow
TLDR: A research project managed to run a gigantic AI on one Apple laptop, but it’s so slow and storage-hungry that commenters instantly split into two camps: impressed hackers and brutal skeptics. The big debate is whether this is a glimpse of the future or just a very clever way to wait forever.
A tiny research team just pulled off the kind of stunt that makes the internet gasp, clap, and immediately start roasting. They got Kimi K3, a truly enormous AI model, running on a single Apple laptop with 64 GB of memory. The catch? It’s painfully slow: about 16 seconds per word-piece in the best case on an M1 Max, and much worse if you don’t download the full mountain of files first. Translation for normal people: yes, it works on a laptop, but you may have time to make coffee, fold laundry, and reconsider your life choices before it finishes a reply.
And the community absolutely zeroed in on that. The loudest reaction was basically: cool trick, unusable reality. One commenter bluntly summed it up as “under 0.02 tok/s,” while another dragged the whole idea with the devastating question, “what is the point of projects like this?” That set the mood fast: half impressed science fair, half public intervention. The storage requirement also caused side-eye, with people balking at the 2 TB drive demand and asking whether an internal drive is mandatory.
But not everyone came to sneer. There was some genuine hype from antirez, who suggested faster results are coming with newer Apple machines and even teased decent speed using two Mac Studios. And then, because no comment section can resist chaos, someone cut through the seriousness with the perfect meme grenade: “Will it fit on ESP32??” In other words, this wasn’t just a tech demo — it became a full-on comments-section soap opera about ambition, practicality, and whether making a huge AI crawl across a laptop is genius or performance art.
Key Points
- •Deltafin is presented as a research project that runs the 2.8T-parameter Kimi K3 Mixture-of-Experts model on a single Apple Silicon Mac, including a 64 GB laptop configuration.
- •The article reports approximate performance of 16 seconds per token on an M1 Max, while full local mode is listed at about 60 to 76 seconds per token and streaming mode at 3+ minutes per token for uncached data.
- •Installation requires a Python environment, dependency installation, compilation of a fused MXFP4 kernel, and model download via `setup_k3.py`.
- •The project offers two storage modes: `--full` at about 1.7 TB with no network use at inference, and `--stream` at about 215 GB with constant network use and much slower uncached inference.
- •Deltafin supports CLI generation, router trace logging, optional int8 spine conversion, and an OpenAI-compatible local server with standard completion endpoints.