Joe Devon

Crazy what a bit of X virality does to a GitHub project's star count.

BREAKTHROUGH: A full, unmodified 2.78-trillion-parameter Kimi K3 on a consumer laptop by streaming only the activated experts from NVMe. YOU CAN’T RUN KIMI K3 “ON THAT” THEY DECLARED. There are many paths to do it. This is one: Marco Bambini Just Gave Us the Full Kimi K3 on a Laptop Meet Marco Bambini he did something that felt impossible only a day ago. He built WASTE Weight-Aware Streaming Tensor Engine a clean, dependency-free C inference engine that runs the complete, unmodified 2.78-trillion-parameter Kimi K3 model by streaming only the activated experts straight from NVMe. No distillation. No pruning. No cloud. The full open-weight model. We have it running in the lab right now. What Marco Actually Built Kimi K3 is a sparse Mixture-of-Experts system. Only about 4 % of its weights fire on any given token. Marco’s insight was simple and ruthless: the idle experts do not need to live in RAM. They only need to be reachable in time. WASTE keeps the model’s “trunk” (attention, shared components, embeddings) resident in memory — roughly 27 GB on the converted container. The 82,000+ routed experts stay on disk as tightly packed residual vector-quantized records. When the router selects its 16 experts per layer, the engine issues direct, cache-bypassing reads from the internal NVMe and feeds them into a bounded expert cache. The rest of the machine’s RAM becomes working space for that cache. On a 64 GB MacBook Pro with the container on the internal SSD, we are measuring 0.32–0.34 tokens per second at a comfortable memory budget. Prefill sits a little higher. The vision tower works. Logits match the reference implementation to within a few parts in a million. It is the real model. The container itself is 982 GiB after conversion from the original 1.42 TB MXFP4 weights. Minimum RAM floor is just over 29 GB for short context. Push the budget higher and the expert cache hit rate climbs; push too high and you start paging and the speed collapses. The sweet spot on current consumer hardware is clear and measurable. How We Are Testing It We converted the official weights, verified the container, and began systematic runs the same day the engine stabilized. First we confirmed numerical fidelity against the PyTorch reference on short prompts. Then we moved to longer generation, vision inputs, and multi-turn chat using Kimi’s native XTML format. We are measuring wall-clock decode, expert I/O versus compute split, cache hit rates at different RAM budgets, and thermal behavior under sustained load. We are also exercising the OpenAI-compatible server that sits on top of the same C library so we can drop the model into existing agent loops without rewriting anything. Early observations: •Expert I/O dominates the timeline, as expected. On a fast internal NVMe the engine is already near the practical ceiling of the storage subsystem. •The architecture’s sparsity is the entire enabler. A dense model of this size would be dead on arrival for local use. •Context length is currently limited by RAM more than by the model itself. Practical working contexts sit comfortably in the tens of thousands of tokens on 64 GB hardware; the full million-token window will need more memory or smarter KV management. •Thinking tokens are expensive at this speed. Long internal monologues turn into multi-hour runs. For agent work we are already experimenting with tighter control over when full reasoning is requested. We are treating this as a research instrument, not a finished product. Every run teaches us something about expert locality, prefetch opportunities, and how far pure software streaming can push trillion-scale inference on ordinary machines. 1 of 2

The job market cannot be this cooked that kids have to get permanent tattoos to get interviews 😭