What LLMs can an M3 Max Mac (36 GB) run?
An M3 Max Mac (36 GB) gives a model 27 GB usable of memory (about 75% of unified memory is usable by the GPU) and 300 GB/s of bandwidth. Of the 61 open models tracked here, 35 fit at Q4_K_M with an 8,192-token context, each leaving at least 0.5 GB free on the M3 Max Mac (36 GB).
Models that fit on one M3 Max Mac (36 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-20b (14.8 GB, 33–56 tokens/s on the M3 Max Mac (36 GB)) is listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B 37.4B, 4B active | GGUF Q4_K_M | 25.4 GB | 13K | 22–37 |
| Ornith 1.5 35B-A3B 36.0B, 3B active | GGUF Q4_K_M | 23.0 GB | 172K | 42–73 |
| Qwen3.6 35B-A3B 36.0B, 3B active | GGUF Q4_K_M | 23.0 GB | 172K | 42–73 |
| Ornith 1.0 35B 35.1B, 3B active | GGUF Q5_K_M | 26.2 GB | 23K | 37–63 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | GGUF Q5_K_M | 24.4 GB | 38K | 27–45 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | GGUF Q5_K_M | 23.5 GB | 256K (full) | 34–58 |
| Gemma 4 31B 31.3B | GGUF Q5_K_M | 25.2 GB | 23K | 6.8–9.3 |
| GLM-4.7 Flash 31.2B, 3B active | GGUF Q5_K_M | 23.6 GB | 58K | 33–57 |
| Xing 4.0 29B-A4B 31.2B, 4B active | GGUF Q5_K_M | 23.6 GB | 68K | 27–46 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | GGUF Q5_K_M | 23.5 GB | 37K | 27–47 |
| Qwen3 30B-A3B 30.5B, 3.3B active | GGUF Q5_K_M | 23.5 GB | 37K | 27–47 |
| Muse Glimmer 30B 29.8B | GGUF Q6_K | 25.7 GB | 63K | 6.6–9.1 |
| Granite 4.2 30B 29.3B | GGUF Q5_K_M | 24.0 GB | 17K | 7.1–9.8 |
| Qwen3.6 27B 27.8B | GGUF Q6_K | 24.4 GB | 38K | 7.0–9.6 |
| Qwen3.8 27B 27.8B | GGUF Q6_K | 24.4 GB | 38K | 7.0–9.6 |
| Hemmingway-1 27B 27.3B | GGUF Q6_K | 24.0 GB | 44K | 7.1–9.8 |
| Gemma 4 26B-A4B 25.8B, 4B active | GGUF Q6_K | 22.7 GB | 186K | 23–39 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 14.8 GB | 128K (full) | 33–56 |
| Gemma 4 12B 12.0B | FP16 / BF16 | 25.7 GB | 56K | 6.7–9.1 |
| ZDTaichu 5.0 9B 9.8B | FP16 / BF16 | 20.8 GB | 128K (full) | 8.2–11 |
| Ornith 1.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 180K | 8.3–11 |
| Qwen3.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 180K | 8.3–11 |
| Ornith 1.0 9B 9.4B | FP16 / BF16 | 20.1 GB | 195K | 8.5–12 |
| MiMo V2.6 Distill Qwen 9B 9.4B | FP16 / BF16 | 20.1 GB | 195K | 8.5–12 |
| Granite 4.2 8B 8.8B | FP16 / BF16 | 19.9 GB | 46K | 8.6–12 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | FP16 / BF16 | 18.0 GB | 128,000 (full) | 28–47 |
| Qwen3 8B 8.2B | FP16 / BF16 | 18.5 GB | 40K (full) | 9.3–13 |
| Llama 3.1 8B 8.0B | FP16 / BF16 | 18.1 GB | 69K | 9.5–13 |
| Gemma 4 E4B 8.0B | FP16 / BF16 | 17.1 GB | 128K (full) | 10–14 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | FP16 / BF16 | 16.7 GB | 128K (full) | 32–55 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 9.35 GB | 451K | 19–26 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 8.78 GB | 256K (full) | 20–27 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 8.69 GB | 128K (full) | 20–28 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 6.02 GB | 128K (full) | 29–41 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 2.79 GB | 128K (full) | 66–96 |
Raising the GPU memory limit on an M3 Max Mac (36 GB)
macOS lets the GPU wire about 27 GB of this Mac's 36 GB by default. Running
sudo sysctl iogpu.wired_limit_mb=28672 raises that to 28 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 28 GB, 35 of the 61 models fit at Q4_K_M with 8K context, against 35 by default
. K2-Horizon MoVA 36B-A4B, the biggest model that fits already, goes from 13K to 18K of context. If macOS runs short of memory it swaps or kills apps, so close big apps first and raise the limit in steps.
Too big for M3 Max Mac (36 GB) at 8K context
How much memory each of these lacks at Q4_K_M (gpt-oss-120b at MXFP4) with 8K context, one request and 0.5 GB kept free, out of the 27 GB usable here; a bigger machine or a smaller quant is the way in.
- Llama 3.1 70B: 20.5 GB short
- Qwen3-Coder-Next: 23.6 GB short
- AliceAI Foundation 80B-A3B: 24.6 GB short
- gpt-oss-120b: 41.2 GB short
- Nemotron 3 Super 120B-A12B: 50.7 GB short
- Qwen3.5 122B-A10B: 51.7 GB short
- Mistral Medium 3.5 128B: 56.2 GB short
- Qwen3.8 Flash Next: 85.8 GB short
- Step 3.7 Flash 196B-A11B: 99.4 GB short
- MiniMax M2.7: 118 GB short
- DeepSeek V4 Flash: 155 GB short
- DeepSeek V4 Flash 0731: 163 GB short
- MiMo V2.6 Flash: 167 GB short
- IQuest-Q1 320B-A15B: 175 GB short
- GLM-5.3 Flash: 173 GB short
- MiniMax M3: 240 GB short
- DeepSeek V3 / R1: 399 GB short
- DeepSeek V3.2: 399 GB short
- GLM-5.3: 442 GB short
- GLM-5.2: 442 GB short
- DeepSeek V4.1 Flash: 448 GB short
- Hy4 Preview 770B-A49B: 458 GB short
- MiMo V2.6 Pro: 609 GB short
- DeepSeek V4 Pro: 966 GB short
- Qwen3.8 2.4T-A95B: 1,491 GB short
- Kimi K3: 1,697 GB short
Best models for M3 Max Mac (36 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 300 GB/s of bandwidth.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| Ornith 1.5 35B-A3B 36.0B | 32K | 23.5 GB | 3.53 GB | 172K | 34–59 |
| Qwen3.6 35B-A3B 36.0B | 32K | 23.5 GB | 3.53 GB | 172K | 34–59 |
| Ornith 1.0 35B 35.1B | 32K | 22.9 GB | 4.05 GB | 197K | 34–59 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B | 32K | 22.6 GB | 4.38 GB | 64K (full) | 20–33 |
| Nemotron 3 Nano 30B-A3B 31.6B | 32K | 20.3 GB | 6.72 GB | 256K (full) | 37–63 |
llama-server commands for an M3 Max Mac (36 GB)
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the M3 Max Mac (36 GB).
Ornith 1.5 35B-A3B, Q4_K_M with 167K tokens: 17–28 tokens/s on the M3 Max Mac (36 GB)
llama-server -hf bartowski/Ornith-1.5-35B-A3B-GGUF:Q4_K_M -c 171008 -ngl 99Ornith-1.5-35B-A3B-Q4_K_M.gguf, 21.9 GB: 26.5 GB used and 526 MB free of the M3 Max Mac (36 GB)'s 27 GB usable, 17–28 tokens/s once the 167K cache is full.
Qwen3.6 35B-A3B, Q4_K_M with 172K tokens: 16–27 tokens/s on the M3 Max Mac (36 GB)
llama-server -hf ggml-org/Qwen3.6-35B-A3B-GGUF:Q4_K_M -c 176128 -ngl 99Qwen3.6-35B-A3B-Q4_K_M.gguf, 20.4 GB: 26.5 GB used and 534 MB free of the M3 Max Mac (36 GB)'s 27 GB usable, 16–27 tokens/s once the 172K cache is full.
Ornith 1.0 35B, Q4_K_M with 190K tokens: 15–26 tokens/s on the M3 Max Mac (36 GB)
llama-server -hf bartowski/deepreinforce-ai_Ornith-1.0-35B-GGUF:Q4_K_M -c 194560 -ngl 99deepreinforce-ai_Ornith-1.0-35B-Q4_K_M.gguf, 21.4 GB: 26.5 GB used and 515 MB free of the M3 Max Mac (36 GB)'s 27 GB usable, 15–26 tokens/s once the 190K cache is full.
How fast the M3 Max Mac (36 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 300 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. With the same memory, M3 Pro Mac (36 GB), 150 GB/s, writes 49% slower on average; M4 Max Mac (36 GB), 410 GB/s, writes 35% faster on average; M5 Max Mac (36 GB), 460 GB/s, writes 51% faster on average.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| Ornith 1.5 35B-A3B | 42–73 | 34–59 | 172K | 16–27 | Does not fit | 151 | 17–29 s |
| Qwen3.6 35B-A3B | 42–73 | 34–59 | 172K | 16–27 | Does not fit | 151 | 17–29 s |
| Ornith 1.0 35B | 42–73 | 34–59 | 197K | 15–25 | Does not fit | 151 | 17–29 s |
| LLM-jp-4.1 32B-A3B Thinking | 30–51 | 20–33 | 64K | 13–22 | Does not fit | 105 | 30–51 s |
| Nemotron 3 Nano 30B-A3B | 39–67 | 37–63 | 256K | 23–39 | Does not fit | 138 | 16–27 s |
| Gemma 4 31B | 7.8–11 | 7.1–9.8 | 61K | 6.5–8.9 | Does not fit | 14 | 102–140 s |
| GLM-4.7 Flash | 38–64 | 24–41 | 117K | 11–18 | Does not fit | 133 | 24–41 s |
| Xing 4.0 29B-A4B | 31–52 | 22–37 | 137K | 10.0–17 | Does not fit | 107 | 27–45 s |
| Qwen3-Coder 30B-A3B | 31–52 | 17–28 | 68K | 10–17 | Does not fit | 107 | 35–59 s |
| Qwen3 30B-A3B | 31–52 | 17–28 | 40K | 15–25 | Does not fit | 107 | 35–59 s |
M3 Max Mac (36 GB) against the other Macs with 27 GB usable
The same models fit on every Mac with 27 GB usable, so speed is what separates them. By bandwidth the M3 Max Mac (36 GB) ranks 3 of 4 at 300 GB/s. Over the 12 biggest models that fit at Q4_K_M with 8K context, the M3 Pro Mac (36 GB), at 150 GB/s, writes 49% slower, the M4 Max Mac (36 GB), at 410 GB/s, writes 35% faster and the M5 Max Mac (36 GB), at 460 GB/s, writes 51% faster than the M3 Max Mac (36 GB). Each cell below is the other Mac's tokens per second minus this one's, midpoints of the estimated ranges.
| Model | M3 Max Mac (36 GB) tokens/s | M3 Pro Mac (36 GB) | M4 Max Mac (36 GB) | M5 Max Mac (36 GB) |
|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B | 22–37 | −14 | +10 | +15 |
| Ornith 1.5 35B-A3B | 42–73 | −28 | +20 | +29 |
| Qwen3.6 35B-A3B | 42–73 | −28 | +20 | +29 |
| Ornith 1.0 35B | 42–73 | −28 | +20 | +29 |
| LLM-jp-4.1 32B-A3B Thinking | 30–51 | −20 | +14 | +21 |
| Nemotron 3 Nano 30B-A3B | 39–67 | −26 | +18 | +26 |
| Gemma 4 31B | 7.8–11 | −4.6 | +3.4 | +4.9 |
| GLM-4.7 Flash | 38–64 | −25 | +18 | +25 |
| Xing 4.0 29B-A4B | 31–52 | −20 | +15 | +21 |
| Qwen3-Coder 30B-A3B | 31–52 | −20 | +14 | +21 |
| Qwen3 30B-A3B | 31–52 | −20 | +14 | +21 |
| Muse Glimmer 30B | 8.9–12 | −5.3 | +3.8 | +5.6 |
Near misses on M3 Max Mac (36 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 |
|---|---|---|
| Granite 4.2 30B 29.3B | 456 MB | Q4_K_M with 28K context |
| K2-Horizon MoVA 36B-A4B 37.4B | 3.31 GB | Q4_K_M with 13K context |
Run locally or rent an M3 Max Mac (36 GB)?
getdeploying.com lists no on-demand rental of M3 Max Mac (36 GB) (checked ). Among the GPUs this site tracks, the cheapest to rent with at least 27 GB is the RTX 5090 at a median $0.69 an hour, which holds 35 of the 61 models at Q4_K_M with 8K context against 35 here; every $1,000 equals about 1,449 hours of it.
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 35 here.
| Setup | How it relates | Memory | Bandwidth | Models at Q4_K_M |
|---|---|---|---|---|
| M3 Pro Mac (36 GB) | Same memory | 27 GB usable | 150 GB/s | 35 (same) |
| M4 Max Mac (36 GB) | Same memory | 27 GB usable | 410 GB/s | 35 (same) |
| M5 Max Mac (36 GB) | Same memory | 27 GB usable | 460 GB/s | 35 (same) |
| M1 Max or M2 Max Mac (32 GB) | Next size down | 21.3 GB usable | 400 GB/s | 28 (−7) |
| M4 Pro Mac (48 GB) | Next size up | 36 GB usable | 273 GB/s | 35 (same) |
Model numbers read from Hugging Face on .