Rent a GPU

Can the Apple M3 Ultra (256GB) run Llama 4 Scout 17B-16E?

Apple M3 Ultra (256GB) × Llama 4 Scout 17B-16E8K context
YesRuns fully on the GPU
Memory at Q8_0118 GB of 192 GB
Weights 115.8 GBKV cache 1.6 GBRuntime 0.6 GB
QuantQ8_0Near-lossless
Speed~27tokens per second
Max context128Kthe model's full window
  1. FP16
  2. Q8_0
  3. Q6_K
  4. Q5_K_M
  5. Q4_K_M
  6. Q3_K_M
  7. Q2_K

Yes. The Apple M3 Ultra (256GB) can run Llama 4 Scout 17B-16E fully in unified memory at Q8_0 (near-lossless), using about 118 GB of the 192 GB available, at roughly 27 tokens/second.

Try other context lengths

Run it with Ollama

With Ollama installed, one command downloads and starts Llama 4 Scout 17B-16E:

ollama run llama4:scout

Ollama's default download is Q4_K_M (about 68.3 GB here). You have room for Q8_0, which is near-lossless: pick that GGUF in LM Studio or from Hugging Face.

New to this? Step-by-step setup for the Apple M3 Ultra (256GB)

Run it in the cloud

No Apple M3 Ultra (256GB)? Run Llama 4 Scout 17B-16E in the cloud.

Rent a cloud GPU by the hour. No upfront cost, stop any time.

Rent a GPU on RunPodRent a GPU on Vast.ai

Memory needed by quantization

At 8,192 tokens of context. The black line marks the 192 GB the Apple M3 Ultra (256GB) can use. macOS lets the GPU use about 192 GB of the 256 GB unified memory by default.

QuantQualityMemoryFits?Speed
FP16Full quality220.2 GBToo big—
Q8_0Near-lossless118 GBFits~27 tok/s
Q6_KExcellent91.6 GBFits~35 tok/s
Q5_K_MVery good79.7 GBFits~41 tok/s
Q4_K_MGood, most popular68.3 GBFits~48 tok/s
Q3_K_MNoticeably degraded55.5 GBFits~59 tok/s
Q2_KHeavily degraded43.1 GBFits~77 tok/s

Measured speed from real Apple M3 Ultra (256GB) owners

Tokens per second that people measured on their own Apple M3 Ultra (256GB), next to our estimate.

No one has measured Llama 4 Scout 17B-16E on the Apple M3 Ultra (256GB) yet. The speeds above are estimates. Run the benchmark and yours becomes the first real number on this page.

Have an Apple M3 Ultra (256GB)? Add your real speed

One command times Llama 4 Scout 17B-16E in your own Ollama, shows you exactly what it would send, and asks before sending anything. No sign-up, no personal data.

macOS and Linux

curl -fsSL https://gpufit.dev/bench.sh | sh -s -- llama4:scout

Windows (PowerShell)

$env:GPUFIT_MODEL="llama4:scout"; irm https://gpufit.dev/bench.ps1 | iex

How the benchmark works

GPUs that run Llama 4 Scout 17B-16E comfortably

Models that run well on the Apple M3 Ultra (256GB)

Frequently asked questions

Can Apple M3 Ultra (256GB) run Llama 4 Scout 17B-16E?

Yes. The Apple M3 Ultra (256GB) can run Llama 4 Scout 17B-16E fully in unified memory at Q8_0 (near-lossless), using about 118 GB of the 192 GB available, at roughly 27 tokens/second.

How much VRAM does Llama 4 Scout 17B-16E need?

About 68.3 GB at Q4_K_M with 8,192 tokens of context, and about 220.2 GB at full FP16 precision.

How fast is Llama 4 Scout 17B-16E on the Apple M3 Ultra (256GB)?

Roughly 27 tokens per second at Q8_0, estimated from the Apple M3 Ultra (256GB)'s 819 GB/s memory bandwidth.