Rent a GPU

Can the Apple M2 Ultra (192GB) run Phi-3.5 Mini 3.8B?

Apple M2 Ultra (192GB) × Phi-3.5 Mini 3.8B8K context
YesRuns fully on the GPU
Memory at FP1611.5 GB of 144 GB
Weights 7.6 GBKV cache 3.2 GBRuntime 0.6 GB
QuantFP16Full quality
Speed~63tokens 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 M2 Ultra (192GB) can run Phi-3.5 Mini 3.8B fully in unified memory at FP16 (full quality), using about 11.5 GB of the 144 GB available, at roughly 63 tokens/second.

Try other context lengths

Run it with Ollama

With Ollama installed, one command downloads and starts Phi-3.5 Mini 3.8B:

ollama run phi3.5:3.8b

Ollama's default download is Q4_K_M (about 6.1 GB here). You have room for FP16, which is full quality: pick that GGUF in LM Studio or from Hugging Face.

New to this? Step-by-step setup for the Apple M2 Ultra (192GB)

Run it in the cloud

No Apple M2 Ultra (192GB)? Run Phi-3.5 Mini 3.8B 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 144 GB the Apple M2 Ultra (192GB) can use. macOS lets the GPU use about 144 GB of the 192 GB unified memory by default.

QuantQualityMemoryFits?Speed
FP16Full quality11.5 GBFits~63 tok/s
Q8_0Near-lossless7.9 GBFits~118 tok/s
Q6_KExcellent7 GBFits~153 tok/s
Q5_K_MVery good6.5 GBFits~177 tok/s
Q4_K_MGood, most popular6.1 GBFits~207 tok/s
Q3_K_MNoticeably degraded5.7 GBFits~257 tok/s
Q2_KHeavily degraded5.3 GBFits~335 tok/s

Measured speed from real Apple M2 Ultra (192GB) owners

Tokens per second that people measured on their own Apple M2 Ultra (192GB), next to our estimate.

No one has measured Phi-3.5 Mini 3.8B on the Apple M2 Ultra (192GB) yet. The speeds above are estimates. Run the benchmark and yours becomes the first real number on this page.

Have an Apple M2 Ultra (192GB)? Add your real speed

One command times Phi-3.5 Mini 3.8B 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 -- phi3.5:3.8b

Windows (PowerShell)

$env:GPUFIT_MODEL="phi3.5:3.8b"; irm https://gpufit.dev/bench.ps1 | iex

How the benchmark works

GPUs that run Phi-3.5 Mini 3.8B comfortably

Models that run well on the Apple M2 Ultra (192GB)

Frequently asked questions

Can Apple M2 Ultra (192GB) run Phi-3.5 Mini 3.8B?

Yes. The Apple M2 Ultra (192GB) can run Phi-3.5 Mini 3.8B fully in unified memory at FP16 (full quality), using about 11.5 GB of the 144 GB available, at roughly 63 tokens/second.

How much VRAM does Phi-3.5 Mini 3.8B need?

About 6.1 GB at Q4_K_M with 8,192 tokens of context, and about 11.5 GB at full FP16 precision.

How fast is Phi-3.5 Mini 3.8B on the Apple M2 Ultra (192GB)?

Roughly 63 tokens per second at FP16, estimated from the Apple M2 Ultra (192GB)'s 800 GB/s memory bandwidth.