Home › Models › Kimi K3 › Q4 (q4_k_m)
Kimi K3 Q4 (q4_k_m): how much VRAM, honestly
Bottom line: 1566.35 GiB weights, ≈1723.72 GiB resident at 8k context → this does not fit ANY single card or Mac in our catalog — a datacenter-class footprint at Q4, not a 24GB project.
Dataset: modelvram.com open dataset (CC BY 4.0 / MIT, retrieved 2026-10-20); "total @8k" = weights + KV cache + runtime overhead per dataset methodology (formula).
Quant ladder around q4_k_m (GiB)
| quant | weights GiB | total @8k GiB |
|---|---|---|
| fp8 | 2589.01 | 2848.65 |
| q8_0 | 2750.83 | 3026.64 |
| q6_k | 2122.99 | 2336.02 |
| q5_k_m | 1834.96 | 2019.19 |
| q4_k_m | 1566.35 | 1723.72 |
| q3_k_m | 1265.38 | 1392.65 |
| q2_k | 1084.15 | 1193.3 |
KV-cache scaling at q4_k_m
| context | total GiB |
|---|---|
| 4k | 1723.6 |
| 8k | 1723.7 |
| 32k | 1724.4 |
| 128k | 1726.9 |
So what can actually run Kimi K3 at q4_k_m?
Needs ≥1723.7 GiB — no single listed rental instance fits (multi-GPU / offload only). modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending. Re-check 2026-10-22 (us): source table values unchanged since its stated 2026-09-29 check.
Cloud-rate check date: 2026-10-22 (us, re-fetching the cited source table: values unchanged since its own stated check 2026-09-29 — the rates below are 3+ weeks old. Provider pages are JS-rendered; verify visually before spending. Commission rates: unverified, no public data.)
If you only have a 24GB card, run one of these instead
LightOnOCR-3 4B (native, 10.1 GiB)Limite 1B Violetto (native, 2.8 GiB)Ornith 1.5 35B-A3B (MoE) (q4_k_m, 22.9 GiB)Ornith 1.5 9B (native, 20.6 GiB)Qwen3.6 35B-A3B (MoE) (q4_k_m, 22.9 GiB)Qwen3.6 27B (q5_k_m, 21.2 GiB)Qwen3.8 27B (q5_k_m, 21.2 GiB)Qwen3.5 9B (native, 20.6 GiB)GLM-4.7 Flash (q5_k_m, 23.6 GiB)Gemma 4 31B (q4_k_m, 21.9 GiB) — from our dataset, best sensible quant, 8k ctx, RTX 4090-class (24GB). More in the calculator or the Kimi K3 on 24GB page.
FAQ
How much VRAM does Kimi K3 q4_k_m need?
About 1723.72 GiB total at 8k context (1566.35 GiB weights, the rest KV cache + runtime overhead). Source: modelvram.com open dataset (CC BY 4.0, retrieved 2026-10-20); extrapolation formula published on the method page.
Does Kimi K3 q4_k_m fit on one GPU?
No single card or Mac in our catalog (up to 256GB unified) fits 1723.72 GiB at q4_k_m. Realistic options: a multi-GPU/multi-node cluster, an API, or the smaller models listed below.
Is there a tier smaller than q4_k_m?
Within our sourced dataset: q3_k_m = 1392.65 GiB at 8k; q2_k = 1193.3 GiB at 8k. Smaller tiers trade accuracy away; we deliberately do not quantify that loss — test the quality on your own workload before trusting outputs.
More: Kimi K3 full quant table · RTX 4090 fits · RunPod vs Vast.ai