Can I Run Llama 4 on 96 GB system RAM?

作者： Jakub Rusinowski · 最后更新： 2025年4月5日

Yes

Yes — Llama 4 Scout 17B at Q4_K_M needs about 68.2 GB of the 76.8 GB usable on 96 GB system RAM (~8.6 GB spare), at ~6 tok/s (estimated), with room for about 32,768 tokens of context.

Confidence: high · Recommended quantization: Q4_K_M · Estimated speed: ~6 tok/s

Buy This HardwareRyzen AI Max+ 395 Laptop (Strix Halo, up to 128GB) — 96 GB VRAM · 120 W board power Deploy in the Cloud NowNVIDIA A100 80GB on RunPod — from $1.39/hr · rate checked 2026-07

or compare on Vast.ai from $0.77/hr (typical low · varies)

As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.

96 GB system RAM — what it gives a model

Usable memory for models	76.8 GB
Memory bandwidth	90 GB/s

Llama 4 on 96 GB system RAM: memory by quantization

Quant	Memory needed	Fits 76.8 GB?	Max context	Est. speed	Download
F16	220.4 GB	✗ No	—	—	218 GB
Q8_0	118.2 GB	✗ No	—	—	115.8 GB
Q6_K	91.8 GB	✗ No	—	—	89.4 GB
Q5_K_M	79.7 GB	✗ No	—	—	77.3 GB
Q4_K_M	68.2 GB	✓ Yes	32K	~6 tok/s	65.8 GB
Q3_K_M	48.9 GB	✓ Yes	128K	~8.2 tok/s	46.5 GB
Q2_K	38.2 GB	✓ Yes	128K	~10.3 tok/s	35.8 GB

Which Llama 4 sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Llama 4 Maverick 17B	243.9 GB	✗ Too large	—
Llama 4 Scout 17B	68.2 GB	✓ Fits	~6 tok/s

What to watch out for

1 larger variant of Llama 4 does not fit and would need CPU offload or different hardware.
These figures assume CPU-only inference. Any discrete GPU, even an 8 GB one, will be several times faster for models that fit in its VRAM.

Recommended setup

llama.cpp (CPU build) or Ollama — both run without a GPU

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
76.8 GB of the 96 GB is treated as usable for model weights (80% — the rest is the OS and running applications).
DDR5-5600 dual channel at 89.6 GB/s peak. CPU decode is assumed to sustain 35% of that peak, because CPU inference is not purely bandwidth-bound — it also spends real time in compute and thread synchronisation. This figure is an assumption, not a fitted constant: no CPU measurement is in the calibration set.
CPU-only inference: no GPU is assumed. A GPU of any size will beat these figures substantially.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is computed from this model's published attention configuration.

FAQ