Hardware requirements

What hardware do I need to run Llama 3.3 70B Instruct?

70.6B 128K context Llama 3.3 Community

At Q4_K_M the weights are 39.7 GB; with an 8K KV cache and runtime overhead that is 42.8 GB, so it wants a 64 GB card to run with headroom — in practice 64 GB VRAM: Instinct MI210. Every figure below is computed from the model’s published shape against each device’s usable memory; tokens per second are estimates. The method.

Minimum practical 40 GB VRAM

A100 40 GB — Q3_K_M, 35.2 GB of 38.4 GB, ~29 tok/s, up to 18K context.

the smallest card that holds it at all, at Q3_K_M — the lowest quantisation worth running
Recommended 64 GB VRAM

Instinct MI210 — Q4_K_M, 42.8 GB of 62.4 GB, ~25 tok/s, up to 70K context.

runs the recommended quantisation with headroom at 8K context
High quality 141 GB VRAM

H200 SXM — Q8_0, 73.0 GB of 139.4 GB, ~42 tok/s, up to 128K context.

near-lossless Q8_0 with headroom
Long context 141 GB VRAM

H200 SXM — Q4_K_M, 80.3 GB of 139.4 GB, ~73 tok/s at 128K.

128K tokens (or the model’s maximum) at the recommended quantisation
Apple Silicon 96 GB unified

M2 Max · 96 GB — Q4_K_M, 42.8 GB of 72.0 GB, ~5.6 tok/s, up to 101K context.

smallest Mac that runs the recommended quantisation with headroom
CPU only 64 GB RAM

CPU only · DDR5 dual-channel — Q5_K_M, 49.7 GB of 51.2 GB, ~0.8 tok/s, up to 12K context.

no GPU — weights stream from system RAM

How much VRAM at each quantisation

QuantWeights+KV 8K+KV 32K+KV 128KTotal @ 8KNeedsQuality
Q8_069.9 GB2.50 GB10.00 GB40.0 GB 73.0 GB141 GB card−0.1% ppl
Q6_K53.9 GB2.50 GB10.00 GB40.0 GB 57.0 GB80 GB card−0.4% ppl
Q5_K_M46.6 GB2.50 GB10.00 GB40.0 GB 49.7 GB64 GB card−0.8% ppl
Q4_K_M39.7 GB2.50 GB10.00 GB40.0 GB 42.8 GB64 GB card−1.9% ppl
Q3_K_M32.1 GB2.50 GB10.00 GB40.0 GB 35.2 GB48 GB card−5.4% ppl
Q2_K27.5 GB2.50 GB10.00 GB40.0 GB 30.6 GB40 GB card−15% ppl

Totals add 0.6 GB runtime overhead and an f16 KV cache; "needs" is the smallest discrete card class whose usable memory (after display and driver reserve) leaves 15% headroom. A q8_0 cache roughly halves the KV columns.

Every device, checked

System RAM for offload:
163264128256

NVIDIA · GeForce

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
GeForce RTX 5090 32 GB — offload — ~6.3 no Check
GeForce RTX 5080 16 GB — no — — no Check
GeForce RTX 5070 Ti 16 GB — no — — no Check
GeForce RTX 5070 12 GB — no — — no Check
GeForce RTX 5060 Ti 16 GB 16 GB — no — — no Check
GeForce RTX 5060 8 GB — no — — no Check
GeForce RTX 4090 24 GB — offload — ~2.6 no Check
GeForce RTX 4080 Super 16 GB — no — — no Check
GeForce RTX 4080 16 GB — no — — no Check
GeForce RTX 4070 Ti Super 16 GB — no — — no Check
GeForce RTX 4070 Ti 12 GB — no — — no Check
GeForce RTX 4070 Super 12 GB — no — — no Check
GeForce RTX 4070 12 GB — no — — no Check
GeForce RTX 4060 Ti 16 GB 16 GB — no — — no Check
GeForce RTX 4060 Ti 8 GB — no — — no Check
GeForce RTX 4060 8 GB — no — — no Check
GeForce RTX 3090 Ti 24 GB — offload — ~2.6 no Check
GeForce RTX 3090 24 GB — offload — ~2.6 no Check
GeForce RTX 3080 Ti 12 GB — no — — no Check
GeForce RTX 3080 12 GB 12 GB — no — — no Check
GeForce RTX 3080 10 GB — no — — no Check
GeForce RTX 3070 Ti 8 GB — no — — no Check
GeForce RTX 3070 8 GB — no — — no Check
GeForce RTX 3060 Ti 8 GB — no — — no Check
GeForce RTX 3060 12 GB 12 GB — no — — no Check
GeForce RTX 2080 Ti 11 GB — no — — no Check
GeForce RTX 2060 12 GB 12 GB — no — — no Check
GeForce GTX 1080 Ti 11 GB — no — — no Check

NVIDIA · GeForce laptop

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
GeForce RTX 5090 Laptop 24 GB — offload — ~2.6 no Check
GeForce RTX 4090 Laptop 16 GB — no — — no Check
GeForce RTX 4080 Laptop 12 GB — no — — no Check
GeForce RTX 4070 Laptop 8 GB — no — — no Check

NVIDIA · Workstation

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
RTX 6000 Ada 48 GB Q3_K_M Q4_K_M 19K 15 no Check
RTX 5000 Ada 32 GB — offload — ~4.7 no Check
RTX 4000 Ada 20 GB — no — — no Check
RTX 2000 Ada 16 GB — no — — no Check
RTX A6000 48 GB Q3_K_M Q4_K_M 19K 12 no Check
RTX A5000 24 GB — offload — ~2.5 no Check
RTX A4000 16 GB — no — — no Check

NVIDIA · Data centre

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
L40S 48 GB Q3_K_M Q4_K_M 19K 13 no Check
L4 24 GB — offload — ~2.2 no Check
A10 24 GB — offload — ~2.5 no Check
A100 40 GB 40 GB — Q3_K_M — — no Check
A100 80 GB 80 GB Q6_K Q8_0 121K 31 no Check
H100 PCIe 80 GB Q6_K Q8_0 121K 30 no Check
H100 SXM 80 GB Q6_K Q8_0 121K 51 no Check
H200 SXM 141 GB Q8_0 Q8_0 128K 73 yes Check
Tesla P40 24 GB — offload — ~2.2 no Check

AMD · Radeon

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
Radeon RX 9070 XT 16 GB — no — — no Check
Radeon RX 9070 16 GB — no — — no Check
Radeon RX 7900 XTX 24 GB — offload — ~2.6 no Check
Radeon RX 7900 XT 20 GB — no — — no Check
Radeon RX 7900 GRE 16 GB — no — — no Check
Radeon RX 7800 XT 16 GB — no — — no Check
Radeon RX 7700 XT 12 GB — no — — no Check
Radeon RX 7600 XT 16 GB — no — — no Check
Radeon RX 6900 XT 16 GB — no — — no Check
Radeon RX 6800 XT 16 GB — no — — no Check
Radeon RX 6700 XT 12 GB — no — — no Check

AMD · Radeon Pro

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
Radeon Pro W7900 48 GB Q3_K_M Q4_K_M 19K 13 no Check
Radeon Pro W7800 32 GB — offload — ~4.7 no Check

AMD · Instinct

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
Instinct MI210 64 GB Q5_K_M Q6_K 70K 25 no Check
Instinct MI300X 192 GB Q8_0 Q8_0 128K 81 yes Check

AMD · Ryzen AI Max

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
Ryzen AI Max+ 395 · 128 GB 128 GB Q8_0 Q8_0 128K 3.6 yes Check
Ryzen AI Max+ 395 · 64 GB 64 GB Q3_K_M Q4_K_M 24K 3.6 no Check

Apple · Apple Silicon

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
M1 · 16 GB 16 GB — no — — no Check
M1 Pro · 32 GB 32 GB — no — — no Check
M1 Max · 64 GB 64 GB Q3_K_M Q4_K_M 24K 5.6 no Check
M1 Ultra · 128 GB 128 GB Q8_0 Q8_0 128K 11 yes Check
M2 · 24 GB 24 GB — no — — no Check
M2 Pro · 32 GB 32 GB — no — — no Check
M2 Max · 96 GB 96 GB Q6_K Q6_K 101K 5.6 no Check
M2 Ultra · 192 GB 192 GB Q8_0 Q8_0 128K 11 yes Check
M3 · 24 GB 24 GB — no — — no Check
M3 Pro · 36 GB 36 GB — no — — no Check
M3 Max · 64 GB 64 GB Q3_K_M Q4_K_M 24K 5.6 no Check
M3 Max · 128 GB 128 GB Q8_0 Q8_0 128K 5.6 yes Check
M3 Ultra · 256 GB 256 GB Q8_0 Q8_0 128K 11 yes Check
M3 Ultra · 512 GB 512 GB Q8_0 Q8_0 128K 11 yes Check
M4 · 32 GB 32 GB — no — — no Check
M4 Pro · 48 GB 48 GB — Q3_K_M — — no Check
M4 Pro · 64 GB 64 GB Q3_K_M Q4_K_M 24K 3.8 no Check
M4 Max · 64 GB 64 GB Q3_K_M Q4_K_M 24K 5.8 no Check
M4 Max · 128 GB 128 GB Q8_0 Q8_0 128K 7.7 yes Check

Intel · Arc

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
Arc B580 12 GB — no — — no Check
Arc B570 10 GB — no — — no Check
Arc A770 16 GB 16 GB — no — — no Check
Arc A750 8 GB — no — — no Check

CPU · CPU + system RAM

DeviceMemoryBest quant with headroomFits at allMax context @ Q4_K_MTok/s est.128K
CPU only · DDR4 dual-channel 32 GB — no — — no Check
CPU only · DDR5 dual-channel 64 GB Q4_K_M Q5_K_M 34K 1.0 no Check
CPU only · DDR5 8-channel server 256 GB Q8_0 Q8_0 128K 3.2 yes Check