LLM Roofline
Apple · UMA · Backend: mlx

Apple Mac Studio M2 Ultra (192 GB) — LLM Sizing & Speed Matrix

800 GB/s dual-die Ultra interconnect; widely available used/refurbished for 120B-235B MoE.

Reference Pricing
$4,999 Street
Usable Memory
149.76 GiB (168.96 tuned)
Memory Bandwidth
800 GB/s (78% eff.)
FP16 Compute
54.4 TFLOPS
Roofline Ridge
48.8 FLOP/B
Power (Load / Idle)
215W / 14W

Buy or Rent Apple Mac Studio M2 Ultra (192 GB)

Vendor & Cloud Links
Mac Studio Ultra at B&H Photo
B&H Photo · Direct Hardware
↗
Apple Mac Studio on Amazon
Amazon · Direct Hardware
↗

All 32 Open-Weight Models on Apple Mac Studio M2 Ultra (192 GB) (Q4_K_M @ 32k)

Model Status Weights Total @ 32k Decode TTFT (4k) Max Safe Ctx
Mixtral 8x22B Instruct COMFORTABLE 81.2 GiB 89.6 GiB 17.4 tok/s 10.90s 64k
Command A 111B COMFORTABLE 63.9 GiB 68.9 GiB 8.7 tok/s 30.75s 256k
Llama 4 Scout 109B-A17B COMFORTABLE 62.8 GiB 70.0 GiB 32.4 tok/s 4.88s 560k
Qwen2.5 72B Instruct COMFORTABLE 41.9 GiB 53.5 GiB 11.2 tok/s 20.31s 32k
DeepSeek-R1 Distill Llama 70B COMFORTABLE 40.7 GiB 52.3 GiB 11.5 tok/s 19.74s 128k
Llama 3.3 70B Instruct COMFORTABLE 40.7 GiB 52.3 GiB 11.5 tok/s 19.74s 128k
Llama 3.1 70B Instruct COMFORTABLE 40.7 GiB 52.3 GiB 11.5 tok/s 19.74s 128k
Llama-3.3 Nemotron Super 49B COMFORTABLE 28.7 GiB 36.5 GiB 16.7 tok/s 13.90s 128k
Mixtral 8x7B Instruct COMFORTABLE 26.9 GiB 32.0 GiB 44.8 tok/s 3.65s 32k
Qwen3 32B COMFORTABLE 18.9 GiB 28.1 GiB 21.6 tok/s 9.43s 40k
Qwen2.5 Coder 32B COMFORTABLE 18.9 GiB 28.1 GiB 21.6 tok/s 9.22s 32k
DeepSeek-R1 Distill Qwen 32B COMFORTABLE 18.9 GiB 28.1 GiB 21.6 tok/s 9.22s 128k
QwQ 32B Reasoning COMFORTABLE 18.9 GiB 28.1 GiB 21.6 tok/s 9.22s 40k
GLM-4 32B 0414 COMFORTABLE 18.6 GiB 21.9 GiB 28.3 tok/s 9.13s 32k
Qwen3 30B-A3B COMFORTABLE 17.6 GiB 21.4 GiB 104.4 tok/s 1.14s 40k
Gemma 3 27B IT COMFORTABLE 15.8 GiB 20.0 GiB 31.1 tok/s 7.68s 128k
Mistral Small 3.1 24B COMFORTABLE 13.8 GiB 20.1 GiB 30.9 tok/s 6.67s 128k
Codestral 22B COMFORTABLE 12.8 GiB 21.2 GiB 29.4 tok/s 6.38s 32k
Qwen3 14B COMFORTABLE 8.5 GiB 14.8 GiB 43.0 tok/s 4.24s 40k
Qwen2.5 14B Instruct COMFORTABLE 8.5 GiB 15.8 GiB 40.0 tok/s 4.29s 32k
DeepSeek-R1 Distill Qwen 14B COMFORTABLE 8.5 GiB 15.8 GiB 40.0 tok/s 4.29s 128k
Phi-4 14B COMFORTABLE 8.5 GiB 12.8 GiB 50.1 tok/s 4.21s 16k
Gemma 3 12B IT COMFORTABLE 7.0 GiB 10.4 GiB 62.2 tok/s 3.53s 128k
Qwen3 8B COMFORTABLE 4.7 GiB 10.3 GiB 63.0 tok/s 2.40s 40k
Llama 3.1 8B Instruct COMFORTABLE 4.6 GiB 9.7 GiB 67.4 tok/s 2.34s 128k
Qwen2.5 7B Instruct COMFORTABLE 4.4 GiB 7.2 GiB 94.8 tok/s 2.19s 32k
Gemma 3 4B IT COMFORTABLE 2.5 GiB 4.2 GiB 174.2 tok/s 1.27s 128k
Qwen3 235B-A22B TIGHT 135.4 GiB 142.4 GiB 27.6 tok/s 6.80s 40k
DeepSeek-R1 671B OOM 386.4 GiB 390.1 GiB OOM — OOM
DeepSeek-V3 671B OOM 386.4 GiB 390.1 GiB OOM — OOM
Llama 3.1 405B Instruct OOM 233.8 GiB 252.3 GiB OOM — OOM
Llama 4 Maverick 400B-A17B OOM 231.5 GiB 238.8 GiB OOM — OOM

Interactive Apple Mac Studio M2 Ultra (192 GB) Simulator

141B (39B active) · GQA
800 GB/s · 54.4 TFLOPS
3. Weight Quantization 4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 64k
Memory Allocation Footprint 89.60 / 149.76 GiB usable
Weights: 81.20 GiB KV Cache: 7.00 GiB Overhead: 1.39 GiB Headroom: 60.16 GiB
Decode Speed
17.4
tok/s (batch=1)
Prefill Speed
377
prompt tok/s
TTFT Latency
10.90s
for 4,096 tok
Active Params
39.0B
of 141B total
llama.cpp Server Command 
llama-server -hf mistralai/Mixtral-8x22B-Instruct-v0.1:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080
Context KV Cache Total VRAM Decode Status
4k 0.88 GiB 83.05 GiB 21.9 tok/s COMFORTABLE
8k 1.75 GiB 84.35 GiB 21.1 tok/s COMFORTABLE
16k 3.50 GiB 86.10 GiB 19.7 tok/s COMFORTABLE
32k 7.00 GiB 89.60 GiB 17.4 tok/s COMFORTABLE
64k 14.00 GiB 96.60 GiB 14.0 tok/s COMFORTABLE

Buy-vs-Cloud-vs-API Monthly TCO Break-Even

BUY LOCAL
Own Hardware/mo
$102.12
276.9h active + power
Cloud GPU/mo
$1111.18
@$3.49/hr spot/pod
Serverless API/mo
$190.00
Payback: 18.2 mo

Recommended Hardware & Cloud Deployment

Contextual to fit status

Hardware FAQ — Apple Mac Studio M2 Ultra (192 GB)

How much usable VRAM / GPU memory does Apple Mac Studio M2 Ultra (192 GB) actually provide for LLMs?

Out of the box, Apple Mac Studio M2 Ultra (192 GB) safely reserves 149.76 GiB for GPU inference, expandable up to 168.96 GiB with OS UMA tuning.

What is the largest Q4_K_M model Apple Mac Studio M2 Ultra (192 GB) can run at 32k context?

With Q4_K_M weights and FP16 KV cache at 32k context, the largest fitting model in our verified catalog on Apple Mac Studio M2 Ultra (192 GB) is Mixtral 8x22B Instruct (141B total).