16GB VRAM

Mac M4 16G — what LLMs can it run?

76 of 145 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.

The short answer for Mac M4 16G

16 GB, 120 GB/s. 76 of 145 models in this index fit comfortably at 4K context; 78 load at all.

Biggest that fits
Hy-Memory-30B-A3B — 9.9 GB at Q8_0-expsIQ2_XXS — 2.1 GB spare at 4K, so longer context comes out of a thin margin.Download: tencent/Hy-Memory-30B-A3B-GGUF · Hy-Memory-30B-A3B-Q8_0-expsIQ2_XXS.gguf · 9.14 GB · listing read 2026-11-12
Room to grow
TwIL 8B balanced — 6.3 GB at Q4_K_M — under 60% of the card, which leaves 5.7 GB for a long context window or a second process.Download: webAI-Official/TwIL-8B · TwIL-8B-Q4_K_M-balanced.gguf · 5.35 GB · listing read 2026-11-15
Speed ceiling
Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 120 GB/s puts a hard ceiling near 26 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
The wall you will hit
The GPU's share of unified memory, not the 16 GB on the box. macOS reserves part of the pool for the system and caps what a single process may wire down, so the practical budget is meaningfully below nameplate — the limit is adjustable (`iogpu.wired_limit_mb`) but it is not absent. Bandwidth is 120 GB/s, which is the number that decides tok/s once a model fits.
Best local LLM for Apple silicon

32B · 1

Mac M4 16G — what LLMs can it run? — 32B
ModelQuantEst. VRAMHeadroomtok/s
Hy-Memory-30B-A3B30B-A3BQ8_0-expsIQ2_XXS9.95 GB+2.1 GB≤139Ceiling

14B · 21

Mac M4 16G — what LLMs can it run? — 14B
ModelQuantEst. VRAMHeadroomtok/s
DeepSeek-Coder-V2-Lite Instruct16B-A2.4BQ4_K_M10.08 GB+1.9 GB≤87Ceiling
DeepSeek-V2-Lite Chat16B-A2.4BQ4_K_M10.08 GB+1.9 GB≤87Ceiling
StarCoder2 15B15BQ4_K_M10.19 GB+1.8 GB10–13Derived
Qwen3 14B Instruct14BQ4_K_M10.04 GB+2 GB10–13Derived
Qwen2.5 14B Instruct14BQ4_K_M10.14 GB+1.9 GB10–14Derived
DeepSeek-R1-Distill-Qwen-14B14BQ4_K_M10.14 GB+1.9 GB10–14Derived
Phi-4 14B14BQ4_K_M10.17 GB+1.8 GB10–14Derived
Phi-3 Medium 14B Instruct14BQ4_K_M9.73 GB+2.3 GB11–14Derived
Ministral 3 14B Instruct14BQ4_K_M9.3 GB+2.7 GB11–15Derived
Mistral Nemo 12B Instruct12BQ4_K_M8.42 GB+3.6 GB12–16Derived
Gemma 3 12B IT12BQ5_K_M9.84 GB+2.2 GB11–14Derived
Mellum2.1 12B-A2.5B Thinking12B-A2.5BQ4_K_M8.59 GB+3.4 GB≤76Ceiling
Stable LM 2 12B Chat12BQ4_K_M8.35 GB+3.7 GB12–16Derived
Gemma 4 12B IT12BQ4_K_M8.02 GB+4 GB13–17Derived
Llama 3.2 11B Vision Instruct11BQ4_K_M7.66 GB+4.3 GB14–18Derived
Solar 10.7B Instruct11BQ4_K_M7.6 GB+4.4 GB14–19Derived
Falcon 3 10B Instruct10BQ4_K_M7.28 GB+4.7 GB14–19Derived
Metis-1.6 Logos10B MoEQ6_K9.31 GB+2.7 GB≤57Ceiling
Gemma 2 9B Instruct9BQ4_K_M7.3 GB+4.7 GB16–22Derived
Qwen3.5 9B9BQ8_010.39 GB+1.6 GB9–12Derived
K2 Horizon 7B7BQ6_K8.35 GB+3.7 GB12–16Derived

7B · 31

Mac M4 16G — what LLMs can it run? — 7B
ModelQuantEst. VRAMHeadroomtok/s
GLM-4-9B-Chat9BQ8_010.16 GB+1.8 GB10–13Derived
MiMo-V2.6-Distill-Qwen-9B9BQ8_010.11 GB+1.9 GB10–13Derived
Qwen3-VL 8B Instruct8BQ8_010.39 GB+1.6 GB10–13Derived
TwIL 8B balanced8.8BQ4_K_M6.28 GB+5.7 GB17–23Derived
Granite 4.2 8B8BQ8_010.46 GB+1.5 GB10–13Derived
Ministral 3 8B Instruct8BQ8_010.02 GB+2 GB10–13Derived
LFM2.5 8B-A1B8B-A1.5BQ8_09.46 GB+2.5 GB≤79Ceiling
Qwen2-VL 7B Instruct7BQ4_K_M5.49 GB+6.5 GB18–24Derived
Granite 3.1 8B Instruct8BQ4_K_M5.88 GB+6.1 GB18–24Derived
Qwen3 8B Instruct8BQ6_K7.64 GB+4.4 GB14–18Derived
Llama 3.1 8B Instruct8BQ8_09.47 GB+2.5 GB11–14Derived
Nous Hermes 3 Llama 3.1 8B8BQ4_K_M5.64 GB+6.4 GB19–25Derived
Aya 23 8B8BQ4_K_M5.64 GB+6.4 GB19–25Derived
OpenChat 3.6 8B8BQ4_K_M5.64 GB+6.4 GB19–25Derived
DeepSeek-R1-Distill-Llama-8B8BQ5_K_M6.51 GB+5.5 GB16–21Derived
InternLM2 7B Chat7BQ4_K_M5.45 GB+6.6 GB19–26Derived
Qwen2.5 7B Instruct7BQ6_K6.77 GB+5.2 GB15–19Derived
Qwen2.5-Coder 7B Instruct7BQ4_K_M5.07 GB+6.9 GB20–26Derived
DeepSeek-R1-Distill-Qwen-7B7BQ4_K_M5.07 GB+6.9 GB20–26Derived
Gemma 4 E4B ITE4BQ8_08.68 GB+3.3 GB11–15Derived
Nemotron 3.5 Bolt E4BE4BQ8_08.5 GB+3.5 GB11–15Derived
OLMo 2 7B Instruct7BQ8_010.3 GB+1.7 GB12–16Derived
OLMo 3 7B Instruct7BQ8_010.3 GB+1.7 GB12–16Derived
WizardLM-2 7B7BQ4_K_M5.14 GB+6.9 GB21–28Derived
Mistral 7B Instruct v0.37BQ6_K6.75 GB+5.3 GB15–20Derived
Zephyr 7B Beta7BQ6_K6.75 GB+5.3 GB15–20Derived
K2 Horizon 3.7B3.7BQ8_06.24 GB+5.8 GB≤23Ceiling
Gemma 4 E2B ITE2BQ8_05.2 GB+6.8 GB≤26Ceiling
Qwen3.5 4B4BQ8_04.97 GB+7 GB≤27Ceiling
Gemma 3 4B IT4BQ8_05.05 GB+7 GB≤28Ceiling
Spark-X2.6 4B4BQ8_04.84 GB+7.2 GB≤29Ceiling

≤3B · 23

Mac M4 16G — what LLMs can it run? — ≤3B
ModelQuantEst. VRAMHeadroomtok/s
Qwen3 4B Instruct4BQ6_K4.05 GB+8 GB≤38Ceiling
Phi-4 Mini Instruct3.8BQ8_04.81 GB+7.2 GB≤31Ceiling
Phi-3.5 Mini Instruct3.8BQ8_05.89 GB+6.1 GB≤31Ceiling
Granite 4.2 3B3BQ8_04.41 GB+7.6 GB≤33Ceiling
Ministral 3 3B Instruct3BQ8_04.26 GB+7.7 GB≤35Ceiling
Llama 3.2 3B Instruct3BQ8_04.05 GB+8 GB≤37Ceiling
Qwen2.5 3B Instruct3BQ8_03.59 GB+8.4 GB≤38Ceiling
SmolLM3 3B3BQ8_03.73 GB+8.3 GB≤39Ceiling
Metis-1.6 Praxis3B MoEQ8_03.4 GB+8.6 GB≤106Ceiling
Gemma 2 2B Instruct2BQ8_03.34 GB+8.7 GB≤46Ceiling
Qwen3.5 2B2BQ8_02.23 GB+9.8 GB≤61Ceiling
LFM3 1.7B1.7BQ8_01.99 GB+10 GB≤69Ceiling
LFM3-VL 2B2BQ8_01.99 GB+10 GB≤69Ceiling
Qwen3 1.7B Instruct1.7BQ8_02.39 GB+9.6 GB≤69Ceiling
Qwen2.5 1.5B Instruct1.5BQ8_01.83 GB+10.2 GB≤77Ceiling
Llama 3.2 1B Instruct1BQ8_01.51 GB+10.5 GB≤96Ceiling
LFM2.5 1.2B Instruct1.2BQ8_01.35 GB+10.7 GB≤102Ceiling
K2 Horizon 0.9B0.9BQ8_01.44 GB+10.6 GB≤110Ceiling
Qwen3.5 0.8B0.8BQ8_00.94 GB+11.1 GB≤148Ceiling
BotPsy 0.8B0.75BQ8_00.9 GB+11.1 GB≤156Ceiling
Qwen2.5 0.5B Instruct0.5BQ8_00.6 GB+11.4 GB≤245Ceiling
MNN-Pocket 0.3B Instruct0.36BQ8_00.46 GB+11.5 GB≤324Ceiling
MNN-Pocket 0.3B QAT0.30BQ4_10.26 GB+11.7 GB≤667Ceiling

How this list is built

Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.

A further 2 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.

Measured on this card

No benchmark runs in this index were recorded on a Mac M4 16G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.

Cards with the same budget

What fits is decided by memory, so every 16GB card of this type returns the same list. These pages are not different answers — they differ in throughput, which this index does not measure per card.

The same 16GB in system memory is a different machine — everything on this page fits there, at a speed set by your DIMMs rather than a graphics bus: 16 GB RAM (CPU)

Same list, different speed — play one sentence on both cards: Mac M4 16G vs Mac M3 16G (Qwen3 14B Instruct) · Mac M4 16G vs Mac M2 16G (Qwen3 14B Instruct) · Mac M4 16G vs Mac M1 16G (Qwen3 14B Instruct)

Setup guides for this card

Step-by-step guides written for the Mac M4 16G's runtime or memory size, with the commands and the checks that prove the model is running on the card.

Stepping up

A Mac M3 Pro 18G (18GB) fits 2 more of the indexed models than this card. Mac M3 Pro 18G →

Before buying, hear the speed difference on a model both cards hold: Mac M4 16G vs Mac M3 Pro 18G (Qwen3 14B Instruct) →

Common questions

What is the best local LLM for a Mac M4 16G?

For everyday use, TwIL 8B balanced at Q4_K_M — about 6.3 GB of the card's 16 GB at 4K context, leaving 5.7 GB for a longer window. If you want the largest thing that will load, that is Hy-Memory-30B-A3B at Q8_0-expsIQ2_XXS (9.9 GB, 2.1 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.

Can a Mac M4 16G run North Mini Code 1.0 30B-A3B?

No. Its smallest build here, Q3_K_M, needs about 15.3 GB and this card has 12.0 GB usable — short by 3.3 GB before any context beyond 4K. The largest model this card does clear is Hy-Memory-30B-A3B.

How many tokens per second does a Mac M4 16G do on an 8B model at Q4?

This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 120 GB/s — a ceiling near 26 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.

Mac M4 16G or RTX 5080 for local LLMs?

They do not hold the same set: Mac M4 16G fits 76 and RTX 5080 84 of 145 at 4K — macOS keeps part of the Mac M4 16G's unified memory from the GPU, leaving about 12 GB usable. Speed is the other difference — 120 GB/s against 960 GB/s, a 8.0× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 5080 generates faster on any model both can hold.

76 of 145 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.

This page's figures change when the model or the runtime does.Last updated 2026-11-17 RSS → /feed.xml