What LLMs can an M4 Max Mac (48 GB) run?
An M4 Max Mac (48 GB) gives a model 36 GB usable of memory (about 75% of unified memory is usable by the GPU) and 546 GB/s of bandwidth. Of the 61 open models tracked here, 35 fit at Q4_K_M with an 8,192-token context, and 3 more at a lower precision, each leaving at least 0.5 GB free on the M4 Max Mac (48 GB).
Models that fit on one M4 Max Mac (48 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, 58–100 tokens/s on the M4 Max Mac (48 GB)) is listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| AliceAI Foundation 80B-A3B 81.3B, 3B active | GGUF IQ3_XXS | 35.1 GB | 25K | 97–173 |
| Qwen3-Coder-Next 79.7B, 3B active | GGUF Q2_K | 34.9 GB | 31K | 96–171 |
| Llama 3.1 70B 70.6B | GGUF Q2_K | 33.5 GB | 13K | 9.2–13 |
| K2-Horizon MoVA 36B-A4B 37.4B, 4B active | GGUF Q6_K | 33.6 GB | 16K | 32–54 |
| Ornith 1.5 35B-A3B 36.0B, 3B active | GGUF Q6_K | 30.9 GB | 223K | 57–99 |
| Qwen3.6 35B-A3B 36.0B, 3B active | GGUF Q6_K | 30.9 GB | 223K | 57–99 |
| Ornith 1.0 35B 35.1B, 3B active | GGUF Q6_K | 30.2 GB | 256K (full) | 57–99 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | FP8 / INT8 | 34.0 GB | 30K | 36–61 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | GGUF Q8_0 | 34.9 GB | 97K | 41–70 |
| Gemma 4 31B 31.3B | FP8 / INT8 | 34.5 GB | 19K | 8.9–12 |
| GLM-4.7 Flash 31.2B, 3B active | GGUF Q8_0 | 34.9 GB | 17K | 42–72 |
| Xing 4.0 29B-A4B 31.2B, 4B active | GGUF Q8_0 | 34.9 GB | 20K | 34–57 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | GGUF Q8_0 | 34.6 GB | 16K | 36–61 |
| Qwen3 30B-A3B 30.5B, 3.3B active | GGUF Q8_0 | 34.6 GB | 16K | 36–61 |
| Muse Glimmer 30B 29.8B | GGUF Q8_0 | 33.1 GB | 128K (full) | 9.3–13 |
| Granite 4.2 30B 29.3B | GGUF Q8_0 | 34.6 GB | 11K | 8.9–12 |
| Qwen3.6 27B 27.8B | GGUF Q8_0 | 31.3 GB | 69K | 9.8–14 |
| Qwen3.8 27B 27.8B | GGUF Q8_0 | 31.3 GB | 69K | 9.8–14 |
| Hemmingway-1 27B 27.3B | GGUF Q8_0 | 30.8 GB | 76K | 10–14 |
| Gemma 4 26B-A4B 25.8B, 4B active | GGUF Q8_0 | 29.1 GB | 256K (full) | 33–56 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 14.8 GB | 128K (full) | 58–100 |
| Gemma 4 12B 12.0B | FP16 / BF16 | 25.7 GB | 256K (full) | 12–17 |
| ZDTaichu 5.0 9B 9.8B | FP16 / BF16 | 20.8 GB | 128K (full) | 15–20 |
| Ornith 1.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 256K (full) | 15–21 |
| Qwen3.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 256K (full) | 15–21 |
| Ornith 1.0 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 15–21 |
| MiMo V2.6 Distill Qwen 9B 9.4B | FP16 / BF16 | 20.1 GB | 256K (full) | 15–21 |
| Granite 4.2 8B 8.8B | FP16 / BF16 | 19.9 GB | 98K | 15–21 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | FP16 / BF16 | 18.0 GB | 128,000 (full) | 49–84 |
| Qwen3 8B 8.2B | FP16 / BF16 | 18.5 GB | 40K (full) | 17–23 |
| Llama 3.1 8B 8.0B | FP16 / BF16 | 18.1 GB | 128K (full) | 17–24 |
| Gemma 4 E4B 8.0B | FP16 / BF16 | 17.1 GB | 128K (full) | 18–25 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | FP16 / BF16 | 16.7 GB | 128K (full) | 56–98 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 9.35 GB | 684K | 33–46 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 8.78 GB | 256K (full) | 35–49 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 8.69 GB | 128K (full) | 36–50 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 6.02 GB | 128K (full) | 51–73 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 2.79 GB | 128K (full) | 112–168 |
Raising the GPU memory limit on an M4 Max Mac (48 GB)
macOS lets the GPU wire about 36 GB of this Mac's 48 GB by default. Running
sudo sysctl iogpu.wired_limit_mb=40960 raises that to 40 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 40 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 57K to 76K 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 M4 Max Mac (48 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 36 GB usable here; a bigger machine or a smaller quant is the way in.
- gpt-oss-120b: 32.2 GB short
- Nemotron 3 Super 120B-A12B: 41.7 GB short
- Qwen3.5 122B-A10B: 42.7 GB short
- Mistral Medium 3.5 128B: 47.2 GB short
- Qwen3.8 Flash Next: 76.8 GB short
- Step 3.7 Flash 196B-A11B: 90.4 GB short
- MiniMax M2.7: 109 GB short
- DeepSeek V4 Flash: 146 GB short
- DeepSeek V4 Flash 0731: 154 GB short
- MiMo V2.6 Flash: 158 GB short
- IQuest-Q1 320B-A15B: 166 GB short
- GLM-5.3 Flash: 164 GB short
- MiniMax M3: 231 GB short
- DeepSeek V3 / R1: 390 GB short
- DeepSeek V3.2: 390 GB short
- GLM-5.3: 433 GB short
- GLM-5.2: 433 GB short
- DeepSeek V4.1 Flash: 439 GB short
- Hy4 Preview 770B-A49B: 449 GB short
- MiMo V2.6 Pro: 600 GB short
- DeepSeek V4 Pro: 957 GB short
- Qwen3.8 2.4T-A95B: 1,482 GB short
- Kimi K3: 1,688 GB short
Best models for M4 Max Mac (48 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 546 GB/s of bandwidth.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B 37.4B | 32K | 30.3 GB | 5.69 GB | 57K | 18–30 |
| Ornith 1.5 35B-A3B 36.0B | 32K | 23.5 GB | 12.5 GB | 256K (full) | 60–104 |
| Qwen3.6 35B-A3B 36.0B | 32K | 23.5 GB | 12.5 GB | 256K (full) | 60–104 |
| Ornith 1.0 35B 35.1B | 32K | 22.9 GB | 13.1 GB | 256K (full) | 60–104 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B | 32K | 22.6 GB | 13.4 GB | 64K (full) | 35–59 |
llama-server commands for an M4 Max Mac (48 GB)
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the M4 Max Mac (48 GB). No -ngl: llama.cpp’s -fit, on by default since b7440, places the layers.
K2-Horizon MoVA 36B-A4B, Q4_K_M with 57K tokens: 12–19 tokens/s on the M4 Max Mac (48 GB)
llama-server -hf IFM/K2-Horizon-MoVA-36B-A4B-GGUF:Q4_K_M -c 58368K2-Horizon-MoVA-36B-A4B-Q4_K_M.gguf, 22.4 GB: 35.5 GB used and 549 MB free of the M4 Max Mac (48 GB)'s 36 GB usable, 12–19 tokens/s once the 57K cache is full.
Ornith 1.5 35B-A3B, Q4_K_M with 256K tokens: 22–37 tokens/s on the M4 Max Mac (48 GB)
llama-server -hf bartowski/Ornith-1.5-35B-A3B-GGUF:Q4_K_M -c 262144Ornith-1.5-35B-A3B-Q4_K_M.gguf, 21.9 GB: 28.4 GB used and 7.60 GB free of the M4 Max Mac (48 GB)'s 36 GB usable, 22–37 tokens/s once the 256K cache is full.
How fast the M4 Max Mac (48 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 546 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, M4 Pro Mac (48 GB), 273 GB/s, writes 48% slower on average; M5 Pro Mac (48 GB), 307 GB/s, writes 42% slower on average; M5 Max Mac (48 GB), 614 GB/s, writes 12% faster on average.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B | 38–66 | 18–30 | 57K | 12–19 | Does not fit | 135 | 33–56 s |
| Ornith 1.5 35B-A3B | 74–129 | 60–104 | 256K | 22–37 | Does not fit | 275 | 10–17 s |
| Qwen3.6 35B-A3B | 74–129 | 60–104 | 256K | 22–37 | Does not fit | 275 | 10–17 s |
| Ornith 1.0 35B | 74–129 | 60–104 | 256K | 22–37 | Does not fit | 275 | 10–17 s |
| LLM-jp-4.1 32B-A3B Thinking | 53–91 | 35–59 | 64K | 24–40 | Does not fit | 191 | 17–29 s |
| Nemotron 3 Nano 30B-A3B | 68–118 | 64–111 | 256K | 41–71 | 41–70 | 252 | 9–16 s |
| Gemma 4 31B | 14–19 | 13–18 | 166K | 8.7–12 | Does not fit | 26 | 56–78 s |
| GLM-4.7 Flash | 65–114 | 43–73 | 198K | 13–21 | 42–72 | 242 | 14–23 s |
| Xing 4.0 29B-A4B | 54–93 | 39–67 | 256K | 11–19 | 34–57 | 195 | 15–25 s |
| Qwen3-Coder 30B-A3B | 54–93 | 30–51 | 155K | 9.2–15 | 36–61 | 195 | 20–33 s |
M4 Max Mac (48 GB) against the other Macs with 36 GB usable
The same models fit on every Mac with 36 GB usable, so speed is what separates them. By bandwidth the M4 Max Mac (48 GB) ranks 2 of 4 at 546 GB/s. Over the 12 biggest models that fit at Q4_K_M with 8K context, the M4 Pro Mac (48 GB), at 273 GB/s, writes 48% slower, the M5 Pro Mac (48 GB), at 307 GB/s, writes 42% slower and the M5 Max Mac (48 GB), at 614 GB/s, writes 12% faster than the M4 Max Mac (48 GB). Each cell below is the other Mac's tokens per second minus this one's, midpoints of the estimated ranges.
| Model | M4 Max Mac (48 GB) tokens/s | M4 Pro Mac (48 GB) | M5 Pro Mac (48 GB) | M5 Max Mac (48 GB) |
|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B | 38–66 | −25 | −22 | +6.2 |
| Ornith 1.5 35B-A3B | 74–129 | −48 | −42 | +11 |
| Qwen3.6 35B-A3B | 74–129 | −48 | −42 | +11 |
| Ornith 1.0 35B | 74–129 | −48 | −42 | +11 |
| LLM-jp-4.1 32B-A3B Thinking | 53–91 | −35 | −31 | +8.4 |
| Nemotron 3 Nano 30B-A3B | 68–118 | −45 | −39 | +11 |
| Gemma 4 31B | 14–19 | −8.3 | −7.3 | +2.1 |
| GLM-4.7 Flash | 65–114 | −43 | −38 | +10 |
| Xing 4.0 29B-A4B | 54–93 | −36 | −31 | +8.5 |
| Qwen3-Coder 30B-A3B | 54–93 | −36 | −31 | +8.5 |
| Qwen3 30B-A3B | 54–93 | −36 | −31 | +8.5 |
| Muse Glimmer 30B | 16–22 | −9.5 | −8.3 | +2.3 |
Near misses on M4 Max Mac (48 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 |
|---|---|---|
| Qwen3-Coder-Next 79.7B | 14.7 GB | GGUF IQ3_XXS with 32K, 35.0 GB |
Run locally or rent an M4 Max Mac (48 GB)?
getdeploying.com lists no on-demand rental of M4 Max Mac (48 GB) (checked ). Among the GPUs this site tracks, the cheapest to rent with at least 36 GB is the L40S at a median $1.57 an hour, which holds 36 of the 61 models at Q4_K_M with 8K context against 35 here; every $1,000 equals about 637 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 |
|---|---|---|---|---|
| M4 Pro Mac (48 GB) | Same memory | 36 GB usable | 273 GB/s | 35 (same) |
| M5 Pro Mac (48 GB) | Same memory | 36 GB usable | 307 GB/s | 35 (same) |
| M5 Max Mac (48 GB) | Same memory | 36 GB usable | 614 GB/s | 35 (same) |
| M5 Max Mac (36 GB) | Next size down | 27 GB usable | 460 GB/s | 35 (same) |
| M4 Pro Mac (64 GB) | Next size up | 48 GB usable | 273 GB/s | 36 (+1) |
Model numbers read from Hugging Face on .