July 31, 2026
Disk, drama, and a very slow Rome
Run Kimi K3 using 29 GB of RAM at 0.50 tok/s
A giant AI somehow ran on a laptop, and the comments instantly turned into a brawl
TLDR: A developer proved a giant AI can run locally on a consumer laptop instead of a data center, but it answers at a snail’s pace. The community split instantly between “wow, the future!” and “cool trick, but is this useful—or just wildly inefficient?”
A developer dropped the kind of demo that makes tech people do a double take: a massive AI model—the sort normally associated with giant data centers—was shown answering a question on a regular 64 GB MacBook Pro, with as little as 29 GB of memory needed to open it. The catch? It’s painfully slow. We’re talking roughly one token every few seconds, or nearly a minute just to answer “What is the capital of Italy?” And honestly, that’s exactly why the comment section lit up.
Some readers were impressed that this worked at all, calling it a glimpse of a future where sensitive data can stay on your own machine instead of being sent to an online AI service. Others immediately went for the jugular. One of the sharpest digs was the accusation that the README “hits all my instincts” of being written by AI, which is the kind of insult that lands like a thrown drink at a dinner party. Another commenter zeroed in on the practical absurdity: if this model is famously long-winded, can it even “get any thinking done” before your coffee gets cold?
Then came the efficiency doom-posting. One user compared the laptop’s power use to a modern AI server and basically declared the whole thing an electricity-hungry stunt. Another wanted to know the most relatable thing of all: where do you even download the 1 TB file? Add in the classic “would Apple’s graphics system make it faster?” question, and the mood was clear: part awe, part skepticism, part meme. The demo says it’s possible. The crowd is arguing over whether that means it’s useful.
Key Points
- •WASTE is a C-based inference engine that runs large models by keeping the model trunk in RAM and streaming selected experts from disk.
- •The article’s main demonstration is the full 2.78 trillion-parameter open-weights Kimi K3 model running on a 64 GB MacBook Pro at about 0.32–0.34 tokens per second.
- •Kimi K3 was converted from a published size of 1.42 TB into a 982 GiB container, with a stated minimum RAM requirement of 29.05 GiB for 4K context.
- •The engine is numerically validated against a PyTorch reference, with final logits agreeing to 3.6e-06 and the vision tower to 2.3e-06.
- •The article says the main challenge now is engineering performance, especially reducing bytes read per token and increasing how much of the model can remain in RAM.