What LLMs can an M2 or M3 Mac (24 GB) run?
An M2 or M3 Mac (24 GB) gives a model 16 GB usable of memory (about two thirds of unified memory is usable by the GPU) and 100 GB/s of bandwidth. Of the 61 open models tracked here, 18 fit at Q4_K_M with an 8,192-token context, and 12 more at a lower precision, each leaving at least 0.5 GB free on the M2 or M3 Mac (24 GB).
Models that fit on one M2 or M3 Mac (24 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, 11–19 tokens/s on the M2 or M3 Mac (24 GB)) is listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | GGUF Q2_K | 14.8 GB | 17K | 14–23 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | GGUF Q2_K | 14.1 GB | 225K | 19–32 |
| GLM-4.7 Flash 31.2B, 3B active | GGUF Q2_K | 14.3 GB | 28K | 17–29 |
| Xing 4.0 29B-A4B 31.2B, 4B active | GGUF Q2_K | 14.3 GB | 33K | 14–24 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | GGUF Q2_K | 14.4 GB | 18K | 13–23 |
| Qwen3 30B-A3B 30.5B, 3.3B active | GGUF Q2_K | 14.4 GB | 18K | 13–23 |
| Muse Glimmer 30B 29.8B | GGUF Q2_K | 13.5 GB | 128K (full) | 4.3–5.9 |
| Granite 4.2 30B 29.3B | GGUF Q2_K | 15.3 GB | 8K | 3.8–5.2 |
| Qwen3.6 27B 27.8B | GGUF Q3_K_M | 15.0 GB | 15K | 3.9–5.3 |
| Qwen3.8 27B 27.8B | GGUF Q3_K_M | 15.0 GB | 15K | 3.9–5.3 |
| Hemmingway-1 27B 27.3B | GGUF Q3_K_M | 14.7 GB | 19K | 3.9–5.4 |
| Gemma 4 26B-A4B 25.8B, 4B active | GGUF Q3_K_M | 13.9 GB | 81K | 12–20 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 14.8 GB | 34K | 11–19 |
| Gemma 4 12B 12.0B | GGUF Q8_0 | 14.2 GB | 85K | 4.1–5.6 |
| ZDTaichu 5.0 9B 9.8B | GGUF Q8_0 | 11.4 GB | 126K | 5.1–7.0 |
| Ornith 1.5 9B 9.7B | GGUF Q8_0 | 11.3 GB | 130K | 5.2–7.1 |
| Qwen3.5 9B 9.7B | GGUF Q8_0 | 11.3 GB | 130K | 5.2–7.1 |
| Ornith 1.0 9B 9.4B | GGUF Q8_0 | 11.0 GB | 138K | 5.3–7.3 |
| MiMo V2.6 Distill Qwen 9B 9.4B | GGUF Q8_0 | 11.0 GB | 138K | 5.3–7.3 |
| Granite 4.2 8B 8.8B | GGUF Q8_0 | 11.4 GB | 31K | 5.1–7.0 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | GGUF Q8_0 | 9.82 GB | 128,000 (full) | 17–29 |
| Qwen3 8B 8.2B | GGUF Q8_0 | 10.7 GB | 39K | 5.5–7.5 |
| Llama 3.1 8B 8.0B | GGUF Q8_0 | 10.3 GB | 45K | 5.7–7.8 |
| Gemma 4 E4B 8.0B | GGUF Q8_0 | 9.38 GB | 128K (full) | 6.3–8.6 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | GGUF Q8_0 | 9.15 GB | 128K (full) | 20–34 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 9.35 GB | 166K | 6.3–8.6 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 8.78 GB | 256K (full) | 6.7–9.2 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 8.69 GB | 87K | 6.8–9.3 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 6.02 GB | 128K (full) | 10–14 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 2.79 GB | 128K (full) | 24–33 |
Raising the GPU memory limit on an M2 or M3 Mac (24 GB)
macOS lets the GPU wire about 16 GB of this Mac's 24 GB by default. Running
sudo sysctl iogpu.wired_limit_mb=20480 raises that to 20 GB, leaving
4 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 20 GB, 23 of the 61 models fit at Q4_K_M with 8K context, against 18 by default
: Gemma 4 26B-A4B (10–17 tokens/s), Hemmingway-1 27B (3.2–4.4 tokens/s), Qwen3.6 27B (3.2–4.3 tokens/s), Qwen3.8 27B (3.2–4.3 tokens/s) and Muse Glimmer 30B (3.0–4.1 tokens/s) join the list. gpt-oss-20b, the biggest model that fits already, goes from 34K to 128K 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 M2 or M3 Mac (24 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 16 GB usable here; a bigger machine or a smaller quant is the way in.
- Gemma 4 31B: 6.36 GB short
- Ornith 1.0 35B: 6.93 GB short
- Ornith 1.5 35B-A3B: 7.45 GB short
- Qwen3.6 35B-A3B: 7.45 GB short
- K2-Horizon MoVA 36B-A4B: 9.86 GB short
- Llama 3.1 70B: 31.5 GB short
- Qwen3-Coder-Next: 34.6 GB short
- AliceAI Foundation 80B-A3B: 35.6 GB short
- gpt-oss-120b: 52.2 GB short
- Nemotron 3 Super 120B-A12B: 61.7 GB short
- Qwen3.5 122B-A10B: 62.7 GB short
- Mistral Medium 3.5 128B: 67.2 GB short
- Qwen3.8 Flash Next: 96.8 GB short
- Step 3.7 Flash 196B-A11B: 110 GB short
- MiniMax M2.7: 129 GB short
- DeepSeek V4 Flash: 166 GB short
- DeepSeek V4 Flash 0731: 174 GB short
- MiMo V2.6 Flash: 178 GB short
- IQuest-Q1 320B-A15B: 186 GB short
- GLM-5.3 Flash: 184 GB short
- MiniMax M3: 251 GB short
- DeepSeek V3 / R1: 410 GB short
- DeepSeek V3.2: 410 GB short
- GLM-5.3: 453 GB short
- GLM-5.2: 453 GB short
- DeepSeek V4.1 Flash: 459 GB short
- Hy4 Preview 770B-A49B: 469 GB short
- MiMo V2.6 Pro: 620 GB short
- DeepSeek V4 Pro: 977 GB short
- Qwen3.8 2.4T-A95B: 1,502 GB short
- Kimi K3: 1,708 GB short
Best models for M2 or M3 Mac (24 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 100 GB/s of bandwidth. gpt-oss-20b is at MXFP4, as published.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| gpt-oss-20b 20.9B, MXFP4 | 32K | 15.4 GB | 571 MB | 34K | 9.3–16 |
| Gemma 4 12B 12.0B | 32K | 8.98 GB | 7.02 GB | 256K (full) | 6.6–9.0 |
| ZDTaichu 5.0 9B 9.8B | 32K | 7.67 GB | 8.33 GB | 128K (full) | 7.8–11 |
| Ornith 1.5 9B 9.7B | 32K | 7.58 GB | 8.42 GB | 256K (full) | 7.9–11 |
| Qwen3.5 9B 9.7B | 32K | 7.58 GB | 8.42 GB | 256K (full) | 7.9–11 |
llama-server commands for an M2 or M3 Mac (24 GB)
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the M2 or M3 Mac (24 GB).
gpt-oss-20b, MXFP4 with 34K tokens: 9.1–15 tokens/s on the M2 or M3 Mac (24 GB)
llama-server -hf ggml-org/gpt-oss-20b-GGUF:MXFP4 -c 34816 -ngl 99 -np 1gpt-oss-20b-MXFP4.gguf, 12.1 GB: 15.5 GB used and 518 MB free of the M2 or M3 Mac (24 GB)'s 16 GB usable, 9.1–15 tokens/s once the 34K cache is full. -np 1: one slot, one sliding window.
Gemma 4 12B, Q4_K_M with 256K tokens: 4.5–6.2 tokens/s on the M2 or M3 Mac (24 GB)
llama-server -hf unsloth/gemma-4-12b-it-GGUF:Q4_K_M -c 262144 -ngl 99 -np 1gemma-4-12b-it-Q4_K_M.gguf, 7.1 GB: 12.8 GB used and 3.17 GB free of the M2 or M3 Mac (24 GB)'s 16 GB usable, 4.5–6.2 tokens/s once the 256K cache is full. -np 1: one slot, one sliding window.
How fast the M2 or M3 Mac (24 GB) writes as the context fills
Tokens per second at Q4_K_M (gpt-oss-20b at MXFP4) 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 100 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 Mac (24 GB), 120 GB/s, writes 20% faster on average; M5 Mac (24 GB), 153 GB/s, writes 52% faster on average; M4 Pro Mac (24 GB), 273 GB/s, writes 168% faster on average; M5 Pro Mac (24 GB), 307 GB/s, writes 200% faster on average.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| gpt-oss-20b | 11–19 | 9.3–16 | 34K | 9.1–15 | MXFP4 only | 39 | 64–108 s |
| Gemma 4 12B | 6.9–9.5 | 6.6–9.0 | 256K | 4.5–6.2 | 4.1–5.6 | 13 | 111–152 s |
| ZDTaichu 5.0 9B | 8.8–12 | 7.8–11 | 128K | 5.3–7.3 | 5.1–7.0 | 16 | 94–129 s |
| Ornith 1.5 9B | 8.9–12 | 7.9–11 | 256K | 3.8–5.2 | 5.2–7.1 | 16 | 93–127 s |
| Qwen3.5 9B | 8.9–12 | 7.9–11 | 256K | 3.8–5.2 | 5.2–7.1 | 16 | 93–127 s |
| Ornith 1.0 9B | 9.1–13 | 8.0–11 | 256K | 3.8–5.2 | 5.3–7.3 | 17 | 91–125 s |
| MiMo V2.6 Distill Qwen 9B | 9.1–13 | 8.0–11 | 256K | 3.8–5.2 | 5.3–7.3 | 17 | 91–125 s |
| Granite 4.2 8B | 8.2–11 | 5.1–7.0 | 55K | 3.8–5.1 | 5.1–7.0 | 15 | 143–196 s |
| LFM2.5 8B-A1B | 28–48 | 22–37 | 128,000 | 12–20 | 17–29 | 99 | 27–45 s |
| Qwen3 8B | 8.8–12 | 5.6–7.6 | 40K | 5.0–6.8 | 5.5–7.5 | 16 | 131–179 s |
M2 or M3 Mac (24 GB) against the other Macs with 16 GB usable
The same models fit on every Mac with 16 GB usable, so speed is what separates them. By bandwidth the M2 or M3 Mac (24 GB) is the slowest of the 5 Macs with 16 GB usable at 100 GB/s. Over the 12 biggest models that fit at Q4_K_M with 8K context, the M4 Mac (24 GB), at 120 GB/s, writes 20% faster, the M5 Mac (24 GB), at 153 GB/s, writes 52% faster, the M4 Pro Mac (24 GB), at 273 GB/s, writes 168% faster and the M5 Pro Mac (24 GB), at 307 GB/s, writes 200% faster than the M2 or M3 Mac (24 GB). Each cell below is the other Mac's tokens per second minus this one's, midpoints of the estimated ranges.
| Model | M2 or M3 Mac (24 GB) tokens/s | M4 Mac (24 GB) | M5 Mac (24 GB) | M4 Pro Mac (24 GB) | M5 Pro Mac (24 GB) |
|---|---|---|---|---|---|
| gpt-oss-20b | 11–19 | +3.0 | +7.9 | +26 | +30 |
| Gemma 4 12B | 6.9–9.5 | +1.6 | +4.3 | +14 | +17 |
| ZDTaichu 5.0 9B | 8.8–12 | +2.1 | +5.4 | +18 | +21 |
| Ornith 1.5 9B | 8.9–12 | +2.1 | +5.5 | +18 | +21 |
| Qwen3.5 9B | 8.9–12 | +2.1 | +5.5 | +18 | +21 |
| Ornith 1.0 9B | 9.1–13 | +2.1 | +5.6 | +18 | +22 |
| MiMo V2.6 Distill Qwen 9B | 9.1–13 | +2.1 | +5.6 | +18 | +22 |
| Granite 4.2 8B | 8.2–11 | +1.9 | +5.1 | +16 | +20 |
| LFM2.5 8B-A1B | 28–48 | +7.4 | +19 | +61 | +72 |
| Qwen3 8B | 8.8–12 | +2.1 | +5.5 | +18 | +21 |
| Llama 3.1 8B | 9.1–13 | +2.1 | +5.7 | +18 | +22 |
| Gemma 4 E4B | 11–15 | +2.5 | +6.7 | +22 | +26 |
Near misses on M2 or M3 Mac (24 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 |
|---|---|---|
| Gemma 4 26B-A4B 25.8B | 1.50 GB | GGUF Q3_K_M with 32K, 14.4 GB |
| Muse Glimmer 30B 29.8B | 3.51 GB | GGUF IQ3_XXS with 32K, 13.6 GB |
| Hemmingway-1 27B 27.3B | 3.63 GB | GGUF IQ3_XXS with 32K, 14.2 GB |
| Qwen3.6 27B 27.8B | 3.92 GB | GGUF IQ3_XXS with 32K, 14.4 GB |
| Qwen3.8 27B 27.8B | 3.92 GB | GGUF IQ3_XXS with 32K, 14.4 GB |
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 18 here.
| Setup | How it relates | Memory | Bandwidth | Models at Q4_K_M |
|---|---|---|---|---|
| M4 Mac (24 GB) | Same memory | 16 GB usable | 120 GB/s | 18 (same) |
| M5 Mac (24 GB) | Same memory | 16 GB usable | 153 GB/s | 18 (same) |
| M4 Pro Mac (24 GB) | Same memory | 16 GB usable | 273 GB/s | 18 (same) |
| M5 Pro Mac (24 GB) | Same memory | 16 GB usable | 307 GB/s | 18 (same) |
| M3 Pro Mac (18 GB) | Next size down | 12 GB usable | 150 GB/s | 17 (−1) |
| M4 Mac (32 GB) | Next size up | 21.3 GB usable | 120 GB/s | 28 (+10) |
Model numbers read from Hugging Face on .