Moonshot's 2.8T-Parameter Kimi K3 Runs on a GPU-less Mini PC
A 2.8T-parameter open-weight Kimi K3 model ran on a GPU-less mini PC using disk-streamed MoE experts and native MXFP4 quantization.
Moonshot released Kimi K3 yesterday, a 2.8-trillion-parameter open-weight model likely the largest open release to date. A developer got it running and producing correct answers on a mini PC with no GPU at all, using a Ryzen AI 9 HX 370 and 128GB of RAM. The 1.56TB checkpoint, split across 96 shards, can't fit in RAM on any normal machine, but because K3 is a mixture-of-experts model, only a handful of its 896 experts per layer activate for any given token. The dense components (attention, shared experts) stay quantized in RAM, while routed experts stream from disk on demand, with an LRU cache pinning frequently used ones permanently.
No checkpoint conversion was needed since Moonshot ships K3's experts natively in MXFP4, which the inference engine reads directly from the published safetensors files. Correctness was verified three ways: a bit-exact comparison against Moonshot's reference code, a structural test against the full checkpoint, and a live prompt test that correctly answered a factual question.
Performance is currently unoptimized: model loading takes about 10 minutes, and generating 40 tokens takes roughly 45 minutes, since the MXFP4 kernel is still scalar code with no SIMD. Another model on the same engine (GLM-5.2, 744B) improved from 0.29 to 1.02 words/sec after eight rounds of tuning, but K3 hasn't received that treatment yet. Even so, the demo shows that trillion-scale MoE models can run on consumer hardware without a GPU, given disk-backed expert streaming.
This synthesis was produced from its source by AI; there is no human editor or manual review step. How we work