Rent a GPU

Can the Apple M4 Max (128GB) run Llama 3.3 70B?

Apple M4 Max (128GB) × Llama 3.3 70B8K context
YesRuns fully on the GPU
Memory at Q8_078.3 GB of 96 GB
Weights 75 GBKV cache 2.7 GBRuntime 0.6 GB
QuantQ8_0Near-lossless
Speed~4tokens per second
Max context60Ktokens that fit at Q8_0
  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 M4 Max (128GB) can run Llama 3.3 70B fully in unified memory at Q8_0 (near-lossless), using about 78.3 GB of the 96 GB available, at roughly 4 tokens/second.

Try other context lengths

Run it with Ollama

With Ollama installed, one command downloads and starts Llama 3.3 70B:

ollama run llama3.3:70b

Ollama's default download is Q4_K_M (about 46.1 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 M4 Max (128GB)

Run it in the cloud

No Apple M4 Max (128GB)? Run Llama 3.3 70B 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 96 GB the Apple M4 Max (128GB) can use. macOS lets the GPU use about 96 GB of the 128 GB unified memory by default.

QuantQualityMemoryFits?Speed
FP16Full quality144.5 GBToo big—
Q8_0Near-lossless78.3 GBFits~4 tok/s
Q6_KExcellent61.2 GBFits~6 tok/s
Q5_K_MVery good53.5 GBFits~7 tok/s
Q4_K_MGood, most popular46.1 GBFits~8 tok/s
Q3_K_MNoticeably degraded37.8 GBFits~9 tok/s
Q2_KHeavily degraded29.8 GBFits~12 tok/s

Measured speed from real Apple M4 Max (128GB) owners

Tokens per second that people measured on their own Apple M4 Max (128GB), next to our estimate.

No one has measured Llama 3.3 70B on the Apple M4 Max (128GB) yet. The speeds above are estimates. Run the benchmark and yours becomes the first real number on this page.

Have an Apple M4 Max (128GB)? Add your real speed

One command times Llama 3.3 70B 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 -- llama3.3:70b

Windows (PowerShell)

$env:GPUFIT_MODEL="llama3.3:70b"; irm https://gpufit.dev/bench.ps1 | iex

How the benchmark works

GPUs that run Llama 3.3 70B comfortably

Models that run well on the Apple M4 Max (128GB)

Frequently asked questions

Can Apple M4 Max (128GB) run Llama 3.3 70B?

Yes. The Apple M4 Max (128GB) can run Llama 3.3 70B fully in unified memory at Q8_0 (near-lossless), using about 78.3 GB of the 96 GB available, at roughly 4 tokens/second.

How much VRAM does Llama 3.3 70B need?

About 46.1 GB at Q4_K_M with 8,192 tokens of context, and about 144.5 GB at full FP16 precision.

How fast is Llama 3.3 70B on the Apple M4 Max (128GB)?

Roughly 4 tokens per second at Q8_0, estimated from the Apple M4 Max (128GB)'s 546 GB/s memory bandwidth.