What LLMs can an M3 Ultra Mac Studio (512 GB) run?
An M3 Ultra Mac Studio (512 GB), with 384 GB usable by the GPU, runs 58 of the 69 open models tracked here at Q4_K_M with 8K context; the largest is Ling 3.1 Flash 560B-A25B (572.1B parameters, 355 GB). It reads memory at 819 GB/s; about 75% of unified memory is usable by the GPU. 7 more fit at a lower precision. Each fit leaves at least 0.5 GB free on the M3 Ultra Mac Studio (512 GB) (why a card does not give a model all its memory).
Models that fit on one M3 Ultra Mac Studio (512 GB)
The most precise weights that still fit with 8K tokens of context and one request, the memory that takes, the longest context at that precision and the writing speed. Bigger models come first. gpt-oss-120b (67.7 GB, 69–121 tokens/s on the M3 Ultra Mac Studio (512 GB)) and gpt-oss-20b (14.8 GB, 83–147 tokens/s on the M3 Ultra Mac Studio (512 GB)) are listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| Hy4 Preview 770B-A49B 780.0B, 49B active | GGUF Q2_K | 336 GB | 498K | 11–19 |
| DeepSeek V4.1 Flash 1017 763.2B, 16B active | GGUF Q3_K_M | 383 GB | 9K | 28–47 |
| DeepSeek V4.1 Flash 763.2B, 16B active | GGUF Q3_K_M | 383 GB | 9K | 28–47 |
| GLM-5.3 753.3B, 40B active | GGUF Q3_K_M | 378 GB | 59K | 12–20 |
| GLM-5.2 753.3B, 40B active | GGUF Q3_K_M | 378 GB | 59K | 12–20 |
| DeepSeek V3.2 685.4B, 37B active | GGUF Q3_K_M | 344 GB | 160K (full) | 13–22 |
| DeepSeek V3 / R1 684.5B, 37B active | GGUF Q3_K_M | 344 GB | 160K (full) | 13–22 |
| Ling 3.1 Flash 560B-A25B 572.1B, 25B active | GGUF Q4_K_M | 355 GB | 256K (full) | 16–27 |
| MiniMax M3 427.0B, 23B active | GGUF Q6_K | 360 GB | 188K | 12–20 |
| GLM-5.3 Flash 321.3B, 18B active | GGUF Q8_0 | 350 GB | 1M (full) | 13–21 |
| IQuest-Q1 320B-A15B 320.3B, 15B active | GGUF Q8_0 | 351 GB | 308K | 13–23 |
| MiMo V2.6 Flash 310.8B, 15B active | GGUF Q8_0 | 339 GB | 1M (full) | 15–25 |
| DeepSeek V4 Flash 0731 304.2B, 13B active | GGUF Q8_0 | 332 GB | 561K | 16–28 |
| DeepSeek V4 Flash 290.9B, 13B active | GGUF Q8_0 | 318 GB | 717K | 16–28 |
| MiniMax M2.7 228.7B, 10B active | GGUF Q8_0 | 252 GB | 200K (full) | 19–32 |
| Step 3.7 Flash 196B-A11B 201.4B, 11B active | GGUF Q8_0 | 220 GB | 256K (full) | 20–33 |
| Qwen3.8 Flash Next 180.0B, 6B active | FP16 / BF16 | 370 GB | 256K (full) | 20–33 |
| Mistral Medium 3.5 128B 127.7B | FP16 / BF16 | 265 GB | 256K (full) | 1.7–2.4 |
| Qwen3.5 122B-A10B 125.1B, 10B active | FP16 / BF16 | 257 GB | 256K (full) | 12–20 |
| Nemotron 3.5 Super VL 120B-A12B 124.5B, 12B active | FP16 / BF16 | 256 GB | 1M (full) | 10–17 |
| Nemotron 3 Super 120B-A12B 123.6B, 12B active | FP16 / BF16 | 254 GB | 256K (full) | 10–17 |
| gpt-oss-120b 116.8B, 5.1B active | As published (MXFP4) | 67.7 GB | 128K (full) | 69–121 |
| AliceAI Foundation 80B-A3B 81.3B, 3B active | FP16 / BF16 | 167 GB | 256K (full) | 37–64 |
| Qwen3-Coder-Next 79.7B, 3B active | FP16 / BF16 | 164 GB | 256K (full) | 37–64 |
| Kolibri-1 78.1B, 3.5B active | FP16 / BF16 | 161 GB | 256K (full) | 33–55 |
| Llama 3.1 70B 70.6B | FP16 / BF16 | 148 GB | 128K (full) | 3.1–4.3 |
| K2-Horizon MoVA 36B-A4B 37.4B, 4B active | FP16 / BF16 | 78.9 GB | 512K (full) | 25–42 |
| Ornith 1.5 35B-A3B 36.0B, 3B active | FP16 / BF16 | 74.3 GB | 256K (full) | 38–64 |
| Qwen3.6 35B-A3B 36.0B, 3B active | FP16 / BF16 | 74.3 GB | 256K (full) | 38–64 |
| Ornith 1.0 35B 35.1B, 3B active | FP16 / BF16 | 72.6 GB | 256K (full) | 38–64 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | FP16 / BF16 | 66.9 GB | 64K (full) | 29–49 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | FP16 / BF16 | 65.3 GB | 256K (full) | 33–56 |
| Gemma 4 31B 31.3B | FP16 / BF16 | 66.6 GB | 256K (full) | 6.9–9.5 |
| GLM-4.7 Flash 31.2B, 3B active | FP16 / BF16 | 64.9 GB | 198K (full) | 36–62 |
| Xing 4.0 29B-A4B 31.2B, 4B active | FP16 / BF16 | 64.8 GB | 256K (full) | 28–48 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | FP16 / BF16 | 63.9 GB | 256K (full) | 32–54 |
| Qwen3 30B-A3B 30.5B, 3.3B active | FP16 / BF16 | 63.9 GB | 40K (full) | 32–54 |
| Muse Glimmer 30B 29.8B | FP16 / BF16 | 61.7 GB | 128K (full) | 7.5–10 |
| Granite 4.2 30B 29.3B | FP16 / BF16 | 62.7 GB | 128K (full) | 7.3–10 |
| ThinkingCap Qwen3.8 27B 27.8B | FP16 / BF16 | 58.0 GB | 256K (full) | 7.9–11 |
| Qwen3.6 27B 27.8B | FP16 / BF16 | 58.0 GB | 256K (full) | 7.9–11 |
| Qwen3.8 27B 27.8B | FP16 / BF16 | 58.0 GB | 256K (full) | 7.9–11 |
| Hemmingway-1 27B 27.3B | FP16 / BF16 | 57.0 GB | 256K (full) | 8.1–11 |
| Gemma 4 26B-A4B 25.8B, 4B active | FP16 / BF16 | 53.9 GB | 256K (full) | 28–47 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 14.8 GB | 128K (full) | 83–147 |
| Gemma 4 12B 12.0B | FP16 / BF16 | 25.7 GB | 256K (full) | 18–25 |
| ZDTaichu 5.0 9B 9.8B | FP16 / BF16 | 20.8 GB | 128K (full) | 22–30 |
| Ornith 1.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 256K (full) | 22–31 |
| Qwen3.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 256K (full) | 22–31 |
| LensVLM 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 23–32 |
| Ornith 1.0 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 23–32 |
| MiMo V2.6 Distill Qwen 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 23–32 |
| Bespoke Nimble 9B 9.0B | FP16 / BF16 | 19.1 GB | 8K (full) | 24–33 |
| Granite 4.2 8B 8.8B | FP16 / BF16 | 19.9 GB | 128K (full) | 23–32 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | FP16 / BF16 | 18.0 GB | 128,000 (full) | 71–124 |
| Qwen3 8B 8.2B | FP16 / BF16 | 18.5 GB | 40K (full) | 25–34 |
| Llama 3.1 8B 8.0B | FP16 / BF16 | 18.1 GB | 128K (full) | 25–35 |
| Gemma 4 E4B 8.0B | FP16 / BF16 | 17.1 GB | 128K (full) | 27–37 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | FP16 / BF16 | 16.7 GB | 128K (full) | 81–143 |
| LightOnOCR-3 4B 4.5B | FP16 / BF16 | 10.1 GB | 256K (full) | 45–64 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 9.35 GB | 1M (full) | 48–69 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 8.78 GB | 256K (full) | 51–73 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 8.69 GB | 128K (full) | 52–74 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 6.02 GB | 128K (full) | 74–108 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 2.79 GB | 128K (full) | 155–241 |
Raising the GPU memory limit on an M3 Ultra Mac Studio (512 GB)
macOS lets the GPU wire about 384 GB of this Mac's 512 GB by default. Running
sudo sysctl iogpu.wired_limit_mb=516096 raises that to 504 GB, leaving
8 GB to macOS and the apps next to it, until the next restart
(macOS 14 or newer; iogpu.wired_limit_mb=0 restores the default). LM Studio and Ollama pick the new limit up after a restart of the app.
At 504 GB, 65 of the 69 models fit at Q4_K_M with 8K context, against 58 by default
: DeepSeek V3 / R1 (11–18 tokens/s), DeepSeek V3.2 (11–18 tokens/s), GLM-5.3 (9.7–16 tokens/s), GLM-5.2 (9.7–16 tokens/s), DeepSeek V4.1 Flash 1017 (23–39 tokens/s), DeepSeek V4.1 Flash (23–39 tokens/s) and Hy4 Preview 770B-A49B (8.0–13 tokens/s) join the list. If macOS runs short of memory it swaps or kills apps, so close big apps first and raise the limit in steps.
How much unified memory a Mac's GPU can use, and why.
Too big for M3 Ultra Mac Studio (512 GB) at 8K context
How much memory each of these lacks at Q4_K_M with 8K context, one request and 0.5 GB kept free, out of the 384 GB usable here; a bigger machine or a smaller quant is the way in.
- MiMo V2.6 Pro: 252 GB short
- DeepSeek V4 Pro: 609 GB short
- Qwen3.8 2.4T-A95B: 1,134 GB short
- Kimi K3: 1,340 GB short
Best models for M3 Ultra Mac Studio (512 GB) with room for long context
The biggest models that still fit at Q4_K_M with 32,768 tokens of context (or their whole window, if shorter), the memory left over, the longest context the Mac takes at that precision and the writing speed with that context in the cache, from 819 GB/s of bandwidth.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| Ling 3.1 Flash 560B-A25B 572.1B | 32K | 355 GB | 28.6 GB | 256K (full) | 16–26 |
| MiniMax M3 427.0B | 32K | 269 GB | 115 GB | 917K | 13–23 |
| GLM-5.3 Flash 321.3B | 32K | 200 GB | 184 GB | 1M (full) | 21–35 |
| IQuest-Q1 320B-A15B 320.3B | 32K | 204 GB | 180 GB | 512K (full) | 18–30 |
| MiMo V2.6 Flash 310.8B | 32K | 194 GB | 190 GB | 1M (full) | 24–40 |
llama-server commands for an M3 Ultra Mac Studio (512 GB)
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the M3 Ultra Mac Studio (512 GB). No -ngl: llama.cpp’s -fit, on by default since b7440, places the layers.
MiniMax M3, Q4_K_M with 872K tokens: 2.0–3.3 tokens/s on the M3 Ultra Mac Studio (512 GB)
llama-server -hf unsloth/MiniMax-M3-GGUF:Q4_K_M -c 892928UD-Q4_K_M/
How fast the M3 Ultra Mac Studio (512 GB) writes as the context fills
Tokens per second at Q4_K_M for the biggest models that fit with 32K tokens, with 8K and 32K tokens in the cache and at the longest context the Mac holds, and at Q8_0 with 8K where that fits. Each token reads the active weights and the whole cache once, so the speed follows the 819 GB/s of bandwidth and falls as the cache grows. The ceiling is that bandwidth divided by the bytes read per token, which no runtime reaches; the reply time is for 1,000 new tokens with 32K (or the longest context) in the cache, prompt processing aside.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| Ling 3.1 Flash 560B-A25B | 16–27 | 16–26 | 256K | 13–23 | Does not fit | 54 | 38–64 s |
| MiniMax M3 | 16–27 | 13–23 | 917K | 1.9–3.2 | Does not fit | 55 | 44–75 s |
| GLM-5.3 Flash | 22–36 | 21–35 | 1M | 7.7–13 | 13–21 | 74 | 29–48 s |
| IQuest-Q1 320B-A15B | 21–36 | 18–30 | 512K | 3.8–6.4 | 13–23 | 74 | 34–57 s |
| MiMo V2.6 Flash | 25–43 | 24–40 | 1M | 7.3–12 | 15–25 | 87 | 25–42 s |
| DeepSeek V4 Flash 0731 | 27–47 | 22–37 | 1M | 2.4–4.1 | 16–28 | 95 | 27–45 s |
| DeepSeek V4 Flash | 27–47 | 22–37 | 1M | 2.4–4.1 | 16–28 | 95 | 27–45 s |
| MiniMax M2.7 | 29–49 | 17–28 | 200K | 4.2–7.0 | 19–32 | 101 | 36–60 s |
| Step 3.7 Flash 196B-A11B | 32–55 | 28–48 | 256K | 12–21 | 20–33 | 114 | 21–36 s |
| Qwen3.8 Flash Next | 59–101 | 51–88 | 256K | 24–40 | 35–60 | 214 | 11–20 s |
Near misses on M3 Ultra Mac Studio (512 GB)
Models that miss at Q4_K_M with 32,768 tokens of context (or their whole window), or leave under 0.5 GB free there, but fit with a shorter context or 3-bit weights, or are short by at most a quarter of the memory and fit with MoE offload. Smallest shortfall first.
| Model | Short by at Q4_K_M | Fits instead |
|---|---|---|
| DeepSeek V3 / R1 684.5B | 43.1 GB | GGUF Q3_K_M with 32K, 346 GB |
| DeepSeek V3.2 685.4B | 43.9 GB | GGUF Q3_K_M with 32K, 346 GB |
| GLM-5.3 753.3B | 86.5 GB | GGUF Q3_K_M with 32K, 381 GB |
| GLM-5.2 753.3B | 86.5 GB | GGUF Q3_K_M with 32K, 381 GB |
| DeepSeek V4.1 Flash 1017 763.2B | 92.3 GB | GGUF IQ3_XXS with 32K, 326 GB |
Run locally or rent an M3 Ultra Mac Studio (512 GB)?
getdeploying.com lists no on-demand rental of M3 Ultra Mac Studio (512 GB) (checked ). None of the GPUs this site tracks has 384 GB of memory on one card; rent several cards or compare multi-GPU servers on the same index.
How these numbers are worked out
Memory is the weights at each precision plus the KV cache for 8,192 tokens in FP16 and runtime overhead (0.5 GB plus 10%), computed from each model's files on Hugging Face with the LLM VRAM Calculator. Speed is estimated from memory bandwidth and the parameters read per token with the LLM Speed Calculator. Both tools take any other model from Hugging Face.
Nearby GPUs and setups
How many of the tracked models each one holds at Q4_K_M with 8K context, against 58 here.
| Setup | How it relates | Memory | Bandwidth | Models at Q4_K_M |
|---|---|---|---|---|
| B200 | Next size down | 180 GB | 8,000 GB/s | 51 (−7) |
| 8× H100 SXM | Next size up | 640 GB | 26,800 GB/s | 66 (+8) |
Model numbers read from Hugging Face on .