LLM Roofline
Apple · UMA · Backend: mlx

Apple Mac Studio M3 Ultra (256 GB) — LLM Sizing & Speed Matrix

Runs Qwen3 235B-A22B Q4_K_M + huge KV cache smoothly at ~35 tok/s under ~230W wall draw.

Reference Pricing
$5,599 Street
Usable Memory
199.68 GiB (225.28 tuned)
Memory Bandwidth
819 GB/s (78% eff.)
FP16 Compute
58 TFLOPS
Roofline Ridge
52.7 FLOP/B
Power (Load / Idle)
230W / 15W

Buy or Rent Apple Mac Studio M3 Ultra (256 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 M3 Ultra (256 GB) (Q4_K_M @ 32k)

Model Status Weights Total @ 32k Decode TTFT (4k) Max Safe Ctx
Qwen3 235B-A22B COMFORTABLE 135.4 GiB 142.4 GiB 28.2 tok/s 6.16s 40k
Mixtral 8x22B Instruct COMFORTABLE 81.2 GiB 89.6 GiB 17.8 tok/s 9.88s 64k
Command A 111B COMFORTABLE 63.9 GiB 68.9 GiB 8.9 tok/s 27.85s 256k
Llama 4 Scout 109B-A17B COMFORTABLE 62.8 GiB 70.0 GiB 33.2 tok/s 4.42s 860k
Qwen2.5 72B Instruct COMFORTABLE 41.9 GiB 53.5 GiB 11.5 tok/s 18.39s 32k
DeepSeek-R1 Distill Llama 70B COMFORTABLE 40.7 GiB 52.3 GiB 11.7 tok/s 17.88s 128k
Llama 3.3 70B Instruct COMFORTABLE 40.7 GiB 52.3 GiB 11.7 tok/s 17.88s 128k
Llama 3.1 70B Instruct COMFORTABLE 40.7 GiB 52.3 GiB 11.7 tok/s 17.88s 128k
Llama-3.3 Nemotron Super 49B COMFORTABLE 28.7 GiB 36.5 GiB 17.1 tok/s 12.59s 128k
Mixtral 8x7B Instruct COMFORTABLE 26.9 GiB 32.0 GiB 45.8 tok/s 3.31s 32k
Qwen3 32B COMFORTABLE 18.9 GiB 28.1 GiB 22.1 tok/s 8.54s 40k
Qwen2.5 Coder 32B COMFORTABLE 18.9 GiB 28.1 GiB 22.1 tok/s 8.35s 32k
DeepSeek-R1 Distill Qwen 32B COMFORTABLE 18.9 GiB 28.1 GiB 22.1 tok/s 8.35s 128k
QwQ 32B Reasoning COMFORTABLE 18.9 GiB 28.1 GiB 22.1 tok/s 8.35s 40k
GLM-4 32B 0414 COMFORTABLE 18.6 GiB 21.9 GiB 29.0 tok/s 8.28s 32k
Qwen3 30B-A3B COMFORTABLE 17.6 GiB 21.4 GiB 106.8 tok/s 1.03s 40k
Gemma 3 27B IT COMFORTABLE 15.8 GiB 20.0 GiB 31.8 tok/s 6.96s 128k
Mistral Small 3.1 24B COMFORTABLE 13.8 GiB 20.1 GiB 31.6 tok/s 6.04s 128k
Codestral 22B COMFORTABLE 12.8 GiB 21.2 GiB 30.1 tok/s 5.78s 32k
Qwen3 14B COMFORTABLE 8.5 GiB 14.8 GiB 44.0 tok/s 3.84s 40k
Qwen2.5 14B Instruct COMFORTABLE 8.5 GiB 15.8 GiB 41.0 tok/s 3.88s 32k
DeepSeek-R1 Distill Qwen 14B COMFORTABLE 8.5 GiB 15.8 GiB 41.0 tok/s 3.88s 128k
Phi-4 14B COMFORTABLE 8.5 GiB 12.8 GiB 51.3 tok/s 3.82s 16k
Gemma 3 12B IT COMFORTABLE 7.0 GiB 10.4 GiB 63.7 tok/s 3.20s 128k
Qwen3 8B COMFORTABLE 4.7 GiB 10.3 GiB 64.5 tok/s 2.18s 40k
Llama 3.1 8B Instruct COMFORTABLE 4.6 GiB 9.7 GiB 69.0 tok/s 2.12s 128k
Qwen2.5 7B Instruct COMFORTABLE 4.4 GiB 7.2 GiB 97.1 tok/s 1.99s 32k
Gemma 3 4B IT COMFORTABLE 2.5 GiB 4.2 GiB 178.3 tok/s 1.15s 128k
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 M3 Ultra (256 GB) Simulator

235.1B (22B active) · GQA
819 GB/s · 58 TFLOPS
3. Weight Quantization 4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 40k
Memory Allocation Footprint 142.38 / 199.68 GiB usable
Weights: 135.40 GiB KV Cache: 5.88 GiB Overhead: 1.11 GiB Headroom: 57.30 GiB
Decode Speed
28.2
tok/s (batch=1)
Prefill Speed
669
prompt tok/s
TTFT Latency
6.16s
for 4,096 tok
Active Params
22.0B
of 235.1B total
llama.cpp Server Command 
llama-server -hf Qwen/Qwen3-235B-A22B:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080
Context KV Cache Total VRAM Decode Status
4k 0.73 GiB 136.96 GiB 39.1 tok/s COMFORTABLE
8k 1.47 GiB 137.98 GiB 37.0 tok/s COMFORTABLE
16k 2.94 GiB 139.45 GiB 33.5 tok/s COMFORTABLE
32k 5.88 GiB 142.38 GiB 28.2 tok/s COMFORTABLE

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

USE SERVERLESS API
Own Hardware/mo
$109.58
168.4h active + power
Cloud GPU/mo
$675.74
@$3.49/hr spot/pod
Serverless API/mo
$19.00
Payback: 346.2 mo

Recommended Hardware & Cloud Deployment

Contextual to fit status

Hardware FAQ — Apple Mac Studio M3 Ultra (256 GB)

How much usable VRAM / GPU memory does Apple Mac Studio M3 Ultra (256 GB) actually provide for LLMs?

Out of the box, Apple Mac Studio M3 Ultra (256 GB) safely reserves 199.68 GiB for GPU inference, expandable up to 225.28 GiB with OS UMA tuning.

What is the largest Q4_K_M model Apple Mac Studio M3 Ultra (256 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 M3 Ultra (256 GB) is Qwen3 235B-A22B (235.1B total).