Home › Models › Kimi K3 › on 24GB VRAM
Can Kimi K3 run on 24GB VRAM? — No, at any sourced quant
A 24GB card gives us 24 GiB usable (our conservative rule — why). Kimi K3 needs, at the smallest tier anyone publishes: 1193.3 GiB at 8k context.
Every tier vs your 24GB budget (GiB total @8k, KV included)
| quant | total @8k | on 24GB (24 GiB usable) |
|---|---|---|
| native | 1599.84 | no — over by ≥1575.8 GiB |
| bf16 | 5696.56 | no — over by ≥5672.6 GiB |
| fp8 | 2848.65 | no — over by ≥2824.7 GiB |
| q8_0 | 3026.64 | no — over by ≥3002.6 GiB |
| q6_k | 2336.02 | no — over by ≥2312.0 GiB |
| q5_k_m | 2019.19 | no — over by ≥1995.2 GiB |
| q4_k_m | 1723.72 | no — over by ≥1699.7 GiB |
| q3_k_m | 1392.65 | no — over by ≥1368.7 GiB |
| q2_k | 1193.3 | no — over by ≥1169.3 GiB |
Source: modelvram.com open dataset (CC BY 4.0 / MIT, retrieved 2026-10-20); KV extrapolated linearly per the published formula.
What Kimi K3 actually takes — realistic routes
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.)
- Hosted API if you need this exact model: API prices change month to month, so we cite none here — check the provider's own pricing page the day you decide (no date-stale numbers on this page, on purpose).
- Multi-GPU rental for bursts: compares that keep their gaps visible: RunPod vs Vast.ai, RunPod vs Lambda.
- Buy hardware only if the math wins: break-even calculator.
Runs comfortably on 24GB (same dataset, 8k ctx)
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)Gemma 4 26B-A4B (MoE) (q6_k, 22.7 GiB)Gemma 4 12B (q8_0, 14.2 GiB)
FAQ
Can Kimi K3 run on 24GB VRAM?
No — not at any quant we can source. The smallest tier in the dataset still needs ≈1193.3 GiB at 8k context versus 24 GiB usable on a 24GB card (modelvram.com dataset, CC BY 4.0, retrieved 2026-10-20).
Would CPU/RAM offloading make Kimi K3 fit?
Not in any practical way: weights alone are 1084.15 GiB at the smallest sourced tier — offload would need roughly that much system RAM plus extreme patience. Quantisation shrinks GiB, not the model's category.
What should I run on 24GB instead?
From our dataset (best sensible quant, 8k ctx): LightOnOCR-3 4B, Limite 1B Violetto, Ornith 1.5 35B-A3B (MoE), Ornith 1.5 9B, Qwen3.6 35B-A3B (MoE), Qwen3.6 27B, Qwen3.8 27B, Qwen3.5 9B and more — full list in the calculator.
More: Kimi K3 Q4 deep-dive · full quant table · RTX 4090 page