No — not on this device
Kimi K3 2.8T-A104B at Q4_K_M needs 1564.0 GB against 96.0 GB usable, and the shortfall of 1468.0 GB is more than 128 GB of system RAM can cover at a tolerable speed. A smaller sibling or a lower quantisation is the honest answer here.
The largest open-weight model ever published. 69 of 93 blocks are linear attention, so its cache is tiny; its 2.8T weights are the problem.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 2750.9 GB | 2751.7 GB | — | ~0.6 | −0.1% ppl | 2655.7 GB over |
| Q6_K | 2123.0 GB | 2123.9 GB | — | ~0.8 | −0.4% ppl | 2027.9 GB over |
| Q5_K_M | 1835.0 GB | 1835.8 GB | — | ~1.0 | −0.8% ppl | 1739.8 GB over |
| Q4_K_M | 1563.2 GB | 1564.0 GB | — | ~1.1 | −1.9% ppl | 1468.0 GB over |
| Q3_K_M | 1265.4 GB | 1266.2 GB | — | ~1.4 | −5.4% ppl | 1170.2 GB over |
| Q2_K | 1084.2 GB | 1085.0 GB | — | ~1.6 | −15% ppl | 989.0 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 24 of its 93 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.
How to run it
$ llama-server \
-hf moonshotai/Kimi-K3:Q4_K_M \
-c 8192 -ngl 5
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.