Kimi K3を441GBに枝刈りして、Mac Studio 1台で動かしたA developer pruned Kimi K3 down to 441 GB and ran it on a single Mac Studio…

A developer writing on Zenn has documented running Kimi K3, a large mixture-of-experts language model, on a single Mac Studio after pruning it down to roughly 441 GB. The experiment matters because it suggests that models at the frontier scale, which normally demand multi-GPU server clusters, can be coaxed into running on one desktop workstation when memory footprint is aggressively reduced.

According to the write-up, the author pruned Kimi K3 to about 441 GB so that it would fit within the 512 GB of unified memory on a Mac Studio powered by Apple's M3 Ultra. They then connected Kimi Code CLI as the harness and evaluated the setup against real tasks from SWE-Lancer, solving 5 of 8 problems worth a combined $3,500. Notably, two of those five were tasks that a previous attempt using a 2-bit quantized K2.7 model had failed to solve. The pruned model has been published on Hugging Face under the handle hellohazime.

The distinction between pruning and quantization is worth clarifying, since both are common techniques for shrinking large models. Quantization lowers the numerical precision of weights, for example from 16-bit to 2-bit, which reduces size but can degrade accuracy at very low bit depths. Pruning instead removes parts of the network entirely, such as less important weights or, in the case of mixture-of-experts architectures, entire expert subnetworks that contribute little to the target workload. The author's remark that the pruned K3 solved problems the 2-bit K2.7 could not appears to support the idea that removing components while keeping higher precision on what remains can preserve capability better than crushing precision across the whole model. That comparison is drawn from a small sample, so it should be read as a single data point rather than a general benchmark result.

SWE-Lancer provides useful context for interpreting the outcome. It is a benchmark built around real freelance software engineering jobs, where each task carries a real-world payout value, and success is measured by whether the model's solution passes the associated tests or requirements. The dollar figures therefore reflect the market value assigned to the work rather than any cost of running the model. A score of 5 out of 8, with $3,500 in resolved value, is a limited slice of the full benchmark, but it is a concrete demonstration of an agentic coding loop completing end-to-end tasks locally.

Kimi is the model family from Moonshot AI, and earlier releases such as Kimi K2 have drawn attention for their scale and coding performance. The K-series models use a mixture-of-experts design, in which only a fraction of the total parameters is activated for any given token. That structure is part of what makes pruning attractive, because experts that rarely fire for a particular domain, such as software engineering tasks, can potentially be trimmed with limited impact on that domain. Kimi Code CLI, used here as the execution harness, is the command-line agent tooling associated with the family, giving the model the ability to read files, run commands, and iterate on code.

The Apple Silicon angle is central to why this is feasible on a single box. The M3 Ultra Mac Studio offers up to 512 GB of unified memory shared between CPU and GPU, which lets very large models load without the partitioning across discrete GPUs that a comparable server would require. Local inference frameworks tuned for Apple Silicon, such as MLX and llama.cpp, have made this class of experiment increasingly common among enthusiasts, and this project fits that broader trend of pushing ever larger models onto consumer-accessible hardware.

Several caveats apply. The results come from one developer's blog post rather than an independently reproduced study, the task sample is small, and details of the pruning method and its effect on general performance beyond these tasks are not fully established from the excerpt alone. The published Hugging Face weights, however, make it possible for others to verify the claims. Taken together, the work is best viewed as an encouraging proof of concept that frontier-scale coding models may be brought within reach of a single high-memory workstation.