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 32 GB of system RAM can cover at a tolerable speed. A smaller sibling or a lower quantisation is the honest answer here.

Does not fit 8K context Q4_K_M · 1563.2 GB Modified MIT Released 27 Jul 2026 Vision Not in the Ollama library

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.

Hardware requirements What hardware do I need to run Kimi K3 2.8T-A104B? →

No single GPU or Mac in the catalogue holds it at Q4_K_M — the page says what comes closest. Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

The VRAM budget

weights 1563.2 GB
Weights 1563.2 GB KV cache @ 8K 0.21 GB Runtime overhead 0.6 GB Over budget 1468.0 GB past 96.0 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
Q8_0 2750.9 GB 2751.7 GB — ~0.7 −0.1% ppl 2655.7 GB over
Q6_K 2123.0 GB 2123.9 GB — ~0.9 −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 — ~1.2 −1.9% ppl 1468.0 GB over
Q3_K_M 1265.4 GB 1266.2 GB — ~1.5 −5.4% ppl 1170.2 GB over
Q2_K 1084.2 GB 1085.0 GB — ~1.8 −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

terminal
$ 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.

01Download is 1563.2 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02The GPU shares one pool with the operating system, and these numbers count about 96.0 GB of 128 GB as its share. Under Linux the GPU can usually address more than that, so treat the verdict as conservative.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.
See all models for this rig Compare with another model

Related models on a DGX Spark · 128 GB

ModelTotalTok/s est.
Kimi K2.6 1T-A32B
1027B MoE · same family
Q4_K_M · 578.6 GB — Won't fit
Qwen3.8 2.4T-A95B
2446B MoE · similar size
Q4_K_M · 1376.7 GB — Won't fit
DeepSeek V4 Pro 1.6T-A49B
1650B MoE · similar size
Q4_K_M · 928.9 GB — Won't fit
GLM-5.2 744B-A40B
753B MoE · similar size
Q4_K_M · 424.7 GB — Won't fit
GLM-5.3 744B-A40B
753B MoE · similar size
Q4_K_M · 424.7 GB — Won't fit
DeepSeek-R1 671B
671B MoE · similar size
Q4_K_M · 378.4 GB — Won't fit
DeepSeek V4.1 Flash 552B-A16B
552B MoE · similar size
Q4_K_M · 311.0 GB — Won't fit

Every model against the DGX Spark · 128 GB →

Kimi K3 2.8T-A104B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
1564.0 GB — Won't fit
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
1564.0 GB — Won't fit
M4 · 16 GB
10.7 GB usable · 120 GB/s
1564.0 GB — Won't fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
1564.0 GB — Won't fit
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
1564.0 GB — Won't fit
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
1564.0 GB — Won't fit

All 114 devices, by tier →