Mac › 48 GB
Best local LLMs for a 48GB Mac
Short answer: A 48GB Mac lets the GPU use about 32 GB by default (~67%). That fits Kimi Linear 48B A3B (MoE) at Q4_K_M with room for 8k context. Best picks: Kimi Linear 48B A3B (MoE), Flint 39.8B A2.5B (MoE), Qwen3.6 35B A3B (MoE). Memory figures are estimates. Speeds are measured llama.cpp results where we have them and labeled estimates otherwise.
Checked 2026-10-22 · memory figures are estimates; measured speeds link their source
| Model | Quant | Memory needed (estimated, 8k ctx) | Fits default limit | Speed on Apple M4 Pro (48 GB) | Format |
|---|---|---|---|---|---|
| Kimi Linear 48B A3B (MoE) MoE | Q4_K_M | 32.1 GB | Y | not measured | GGUF |
| Flint 39.8B A2.5B (MoE) MoE | Q5_K_M | 30.6 GB | Y | not measured | GGUF |
| Qwen3.6 35B A3B (MoE) MoE | Q6_K | 31.7 GB | Y | not measured | GGUF |
| Qwen3.5 35B A3B (MoE) MoE | Q6_K | 31.7 GB | Y | not measured | GGUF |
| Olmo 3.1 32B | Q6_K | 29.7 GB | Y | ~7 tok/s est. | GGUF |
| NVIDIA Nemotron 3.5 Lightning 30B A3B (MoE) MoE | Q6_K | 27.8 GB | Y | not measured | GGUF |
| NVIDIA Nemotron 3 Nano 30B A3B (MoE) MoE | Q6_K | 27.8 GB | Y | not measured | GGUF |
| Gemma 4 31B | Q6_K | 29.7 GB | Y | ~7 tok/s est. | GGUF |
| GLM 4.7 Flash MoE | Q6_K | 27.9 GB | Y | not measured | GGUF |
| Qwen3 Coder 30B A3B (MoE) MoE | Q6_K | 27.7 GB | Y | not measured | GGUF |
| Qwen3 30B A3B (2507) (MoE) MoE | Q6_K | 27.7 GB | Y | not measured | GGUF |
| granite 4.2 30b | Q6_K | 28.0 GB | Y | ~8 tok/s est. | GGUF |
| Raon-1 mini preview 28.8B A4.3B (MoE) MoE | Q6_K | 25.4 GB | Y | not measured | GGUF |
| Qwen3.8 27B | Q8_0 | 32.1 GB | Y | ~6 tok/s est. | GGUF |
| Qwen3.6 27B | Q8_0 | 32.1 GB | Y | ~6 tok/s est. | GGUF |
| Qwen3.5 27B | Q8_0 | 32.1 GB | Y | ~6 tok/s est. | GGUF |
| Gemma 4 26B A4B (MoE) MoE | Q8_0 | 29.9 GB | Y | not measured | GGUF |
| Devstral Small 2 24B (2512) | Q8_0 | 28.7 GB | Y | ~7 tok/s est. | GGUF |
| LFM2.5 24B A2B (MoE) MoE | Q8_0 | 27.5 GB | Y | not measured | GGUF |
| gpt-oss 20B (MoE) MoE | Q8_0 | 24.1 GB | Y | not measured | GGUF |
| Phi 4 reasoning vision 15B | Q8_0 | 19.1 GB | Y | ~12 tok/s est. | GGUF |
| Ministral 3 14B (2512) | Q8_0 | 17.5 GB | Y | ~13 tok/s est. | GGUF |
| Gemma 4 12B | Q8_0 | 14.8 GB | Y | ~15 tok/s est. | GGUF |
| GLM 4.6V Flash | Q8_0 | 12.4 GB | Y | ~17 tok/s est. | GGUF |
| Qwen3.5 9B | Q8_0 | 11.6 GB | Y | ~18 tok/s est. | GGUF |
| Ministral 3 8B (2512) | Q8_0 | 11.7 GB | Y | ~20 tok/s est. | GGUF |
| granite 4.2 8b | Q8_0 | 11.8 GB | Y | ~20 tok/s est. | GGUF |
| Qwen3 VL 8B | Q8_0 | 11.6 GB | Y | ~20 tok/s est. | GGUF |
| LFM2.5 8B A1B (MoE) MoE | Q8_0 | 10.1 GB | Y | not measured | GGUF |
| Gemma 4 E4B | Q8_0 | 9.7 GB | Y | ~22 tok/s est. | GGUF |
| Gemma 4 E2B | Q8_0 | 6.4 GB | Y | ~35 tok/s est. | GGUF |
| Qwen3.5 4B | Q8_0 | 6.1 GB | Y | ~38 tok/s est. | GGUF |
| Qwen3 4B (2507) | Q8_0 | 6.3 GB | Y | ~44 tok/s est. | GGUF |
| Nemotron Labs Dory 1.0 4B | Q8_0 | 5.5 GB | Y | ~45 tok/s est. | GGUF |
| NVIDIA Nemotron 3 Nano 4B | Q8_0 | 5.2 GB | Y | ~45 tok/s est. | GGUF |
| Ministral 3 3B (2512) | Q8_0 | 5.8 GB | Y | ~46–47 tok/s est. | GGUF |
| Qwen3.5 2B | Q8_0 | 3.2 GB | Y | ~78–80 tok/s est. | GGUF |
Best quant = highest of Q8_0, Q6_K, Q5_K_M, Q4_K_M, IQ4_XS, Q3_K_M that fits the default limit, f16 KV cache, 8k context. Speeds are llama.cpp (GGUF) figures; we don't have MLX runs. Estimated speeds are scaled from measured Llama 2 7B runs on the same chip and are not done for MoE models.
Why your Mac shows less than 48 GB
macOS caps how much unified memory the GPU can wire. We assume about 67% (about 32 GB here), roughly 67% below 64 GB and 75% at 64 GB and up. You can raise the cap with sudo sysctl iogpu.wired_limit_mb=…, at your own risk; leave room for macOS and your apps. These shares are approximate. Method.
Which Macs have 48GB
Apple M3 Max (48 GB), Apple M4 Pro (48 GB), Apple M4 Max (48 GB), Apple M5 Pro (48 GB), Apple M5 Max (48 GB).
FAQ
Is 48GB enough to run an LLM on a Mac?
Yes. 37 of 64 current models in our list fit fully at Q3_K_M or better with 8k context, the largest being Kimi Linear 48B A3B (MoE) at Q4_K_M (about 32.1 GB, estimate). Bigger models need a Mac with more memory.
How much of my Mac's memory can the GPU use?
By default macOS caps GPU wired memory at roughly 67% of total below 64 GB and roughly 75% at 64 GB and up, so about 32 GB on a 48GB Mac (approximate). The iogpu.wired_limit_mb sysctl can raise it, at your own risk.
Is MLX faster than llama.cpp on a Mac?
We haven't measured it. Public tests conflict: MLX often shows higher generation tok/s, but one M1 Max test found GGUF in llama.cpp finished short-output tasks faster because of prompt processing. See our MLX vs GGUF answer for the sources.
Can a 48GB Mac run a 70B model?
No, not at the default limit. Llama 3.3 70B Instruct needs about 48.2 GB at Q4_K_M with 8k context (estimate), more than the roughly 32 GB the GPU can use.
Mac or a used RTX 3090 for local LLMs?
It's a trade. The 3090 has 24 GB of VRAM and is usually faster on models that fit. A 48GB Mac gives the GPU about 32 GB, so it fits bigger models. Check both in the checker.
Other Mac tiers: 32 GB Mac · 64 GB Mac · all Mac sizes · GPUs: 32 GB VRAM · What to buy
Will It Run? Pro NOT BUILT YET
A personal setup for your exact hardware (Ollama / llama.cpp configs ready for your coding agent), an alert when a new model fits, and a one-click cloud fallback when it doesn't.
Not built yet. Join the waitlist, $10/mo planned, no charge now. What's planned
Email only. No payment info.
Get the weekly "new models & what it takes to run them" email
Free. One email a week. Unsubscribe anytime.