Llama on NVIDIA Ampere

Can I run Llama 4 Scout 109B-A17B on an RTX 3060 12GB?

Not at Q4_K_M — it needs 64.0 GB against 10.5 GB available. You would need 7 of these cards.

Won't fit

64.0 GB of 10.5 GB · 611%
067 GB
Weights 61.7 GB
KV cache 1.5 GB
Runtime overhead 0.8 GB
Over the limit 53.6 GB

Short by 53.6 GB. You can run it with 6 of 48 layers on the RTX 3060 12GB and the rest in system RAM, at roughly 4.27 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation4.27tok/s
Prompt processing309tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of Llama 4 Scout 109B-A17B on a RTX 3060 12GB

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 203.0 GB 205.4 GB Won't fit — 1.25
INT8 / W8A8 8.50 107.4 GB 109.7 GB Won't fit — 2.40
Q8_0 (GGUF) 8.50 107.4 GB 109.7 GB Won't fit — 2.40
FP8 (E4M3) 8.00 101.5 GB 103.9 GB Won't fit — 2.53
Q6_K 6.56 83.2 GB 85.6 GB Won't fit — 3.11
Q5_K_M 5.67 71.9 GB 74.3 GB Won't fit — 3.63
Q5_K_S 5.52 70.0 GB 72.4 GB Won't fit — 3.72
Q4_K_M 4.85 61.7 GB 64.0 GB Won't fit — 4.27
Q4_K_S 4.58 58.3 GB 60.7 GB Won't fit — 4.49
Q4_0 4.55 58.0 GB 60.3 GB Won't fit — 4.52
AWQ 4-bit 4.25 56.8 GB 59.1 GB Won't fit — 4.61
GPTQ 4-bit 4.25 56.8 GB 59.1 GB Won't fit — 4.61
MXFP4 4.25 56.8 GB 59.1 GB Won't fit — 4.61
IQ4_XS 4.25 54.2 GB 56.6 GB Won't fit — 4.89
Q3_K_M 3.91 50.0 GB 52.3 GB Won't fit — 5.27
IQ3_M 3.70 47.4 GB 49.7 GB Won't fit — 5.64
IQ3_XXS 3.06 39.4 GB 41.8 GB Won't fit — 6.78
Q2_K 2.63 34.1 GB 36.4 GB Won't fit — 8.02
IQ2_XXS 2.06 27.0 GB 29.3 GB Won't fit — 10.4
IQ1_M 1.75 23.1 GB 25.4 GB Won't fit — 12.4

Also worth checking

Llama 4 Scout 109B-A17B on every GPU
Full hardware tableEverything that runs on a RTX 3060 12GB
Full model tableWhy the KV cache is the part that bites
Guide