RunsOnMy ModelsGPUs & MacsAPI pricesVRAM calculatorNew

What AI models run on the Mac M4 16GB?

The Mac M4 16GB has 16 GB of unified memory (120 GB/s). About 10.7 GB is usable for a model. 165 of 285 open models we track fit at 8K context.

165models fit (8K ctx)
144still fit at 32K ctx
26Blargest that fits
120 GB/smemory bandwidth

Models that fit, largest first

ModelParamsBest quant (8K)Needs32K contextMax tok/s (est.)
Trinity Mini26B (2.0B act.)Q2_K9.8 GBQ2_K169details
GLiDE 27B26BQ2_K10.5 GBno12details
diffusiongemma 26B A4B it26B (4.0B act.)Q2_K10.4 GBno80details
gemma 4 26B A4B it26B (4.0B act.)Q2_K10.4 GBno80details
LFM2.5 24B A2B24B (2.0B act.)Q2_K9.2 GBQ2_K160details
ERNIE 4.5 21B A3B PT22BQ2_K8.6 GBQ2_K15details
ERNIE 4.5 21B A3B Thinking22BQ2_K8.6 GBQ2_K15details
Groot 19B A3B 26102319B (3.0B act.)Q2_K8.4 GBQ2_K99details
Kimi VL A3B Thinking 250616B (3.0B act.)Q3_K_M8.4 GBQ3_K_M82details
Kimi VL A3B Instruct16B (3.0B act.)Q3_K_M8.7 GBQ3_K_M79details
Kimi VL A3B Thinking16B (3.0B act.)Q3_K_M8.4 GBQ3_K_M82details
deepseek moe 16b base16B (2.8B act.)Q2_K8.6 GBno104details
LLaDA2.1 mini16B (1.4B act.)Q3_K_M8.4 GBQ3_K_M171details
Moonlight 16B A3B Instruct16B (3.0B act.)Q3_K_M8.7 GBQ3_K_M77details
DeepSeek Coder V2 Lite Instruct16B (2.7B act.)Q3_K_M8.0 GBQ3_K_M92details
DeepSeek V2 Lite16B (2.7B act.)Q3_K_M8.0 GBQ3_K_M92details
DeepSeek V2 Lite Chat16B (2.7B act.)Q3_K_M8.4 GBQ3_K_M87details
Qwen2.5 Coder 14B Instruct15BQ3_K_M8.9 GBno17details
DeepSeek R1 Distill Qwen 14B15BQ3_K_M9.0 GBno16details
Qwen3 14B15BQ3_K_M8.7 GBno17details
phi 415BQ3_K_M8.9 GBno17details
gemma 3 12b it12BQ5_K_M9.6 GBQ3_K_M14details
Llama Guard 4 12B12BQ4_K_M9.1 GBno16details
Llama 3.2 11B Vision Instruct11BQ5_K_M9.0 GBQ2_K16details
Llama 3.2 11B Vision11BQ5_K_M9.0 GBQ2_K16details
GLM 4.6V Flash10BQ6_K8.7 GBQ6_K15details
GLM 4.1V 9B Thinking10BQ6_K8.7 GBQ6_K15details
Step3 VL 10B10BQ6_K8.1 GBQ3_K_M18details
SenseNova Loop 9B Preview9.7BQ6_K8.0 GBQ6_K16details
Qwen3.5 9B9.7BQ6_K8.1 GBQ6_K16details
Ornith 1.5 9B9.7BQ6_K8.3 GBQ6_K16details
Qwen3.5 9B Base9.7BQ6_K8.0 GBQ6_K16details
MiMo V2.6 Distill Qwen 9B9.4BQ6_K8.4 GBQ6_K15details
glm 4 9b9.4BQ2_K9.4 GBno30details
internlm3 8b instruct8.8BQ6_K7.8 GBQ6_K17details
granite 4.2 8b8.8BQ6_K8.6 GBQ3_K_M17details
granite 4.1 8b8.8BQ6_K8.6 GBQ3_K_M17details
Qwen3 VL 8B Instruct8.8BQ6_K8.5 GBQ3_K_M17details
Molmo2 8B8.7BQ6_K8.4 GBQ3_K_M17details
LFM2.5 8B A1B8.5B (1.0B act.)Q8_09.1 GBQ8_0113details
Qwen2.5 VL 7B Instruct8.3BQ8_08.7 GBQ6_K15details
UI TARS 1.5 7B8.3BQ8_08.7 GBQ6_K15details
olmOCR 2 7B 10258.3BQ8_08.7 GBQ6_K15details
olmOCR 7B 0225 preview8.3BQ8_08.7 GBQ6_K15details
UI TARS 7B DPO8.3BQ8_08.7 GBQ6_K15details
Qwen3 8B8.2BQ6_K8.1 GBQ3_K_M18details
DeepSeek R1 0528 Qwen3 8B8.2BQ6_K8.1 GBQ3_K_M18details
granite 3.3 8b instruct8.2BQ6_K8.0 GBQ3_K_M18details
granite 3.1 8b instruct8.2BQ6_K8.0 GBQ3_K_M18details
granite 3.2 8b instruct8.2BQ6_K8.0 GBQ3_K_M18details
Apertus 8B Instruct 25098.1BQ8_09.6 GBQ4_K_M14details
Llama 3.1 8B Instruct8.0BQ8_09.6 GBQ4_K_M14details
Llama 3.1 8B8.0BQ8_09.6 GBQ4_K_M14details
Hermes 3 Llama 3.1 8B8.0BQ8_09.6 GBQ4_K_M14details
DeepSeek R1 Distill Llama 8B8.0BQ8_09.6 GBQ4_K_M14details
Meta Llama 3 8B Instruct8.0BQ8_09.6 GBQ4_K_M14details
Meta Llama Guard 2 8B8.0BQ8_09.6 GBQ4_K_M14details
Meta Llama 3.1 8B Instruct8.0BQ8_09.6 GBQ4_K_M14details
Ling 3.0 tiny7.9B (1.4B act.)Q8_08.7 GBQ8_082details
MiMo 7B RL7.8BQ8_09.4 GBQ4_K_M15details
internlm2 5 7b chat7.7BQ8_09.2 GBQ4_K_M15details
internlm2 chat 7b7.7BQ8_09.2 GBQ4_K_M15details
Qwen2.5 7B Instruct7.6BQ8_08.7 GBQ6_K15details
Qwen2.5 Coder 7B Instruct7.6BQ8_08.7 GBQ6_K15details
DeepSeek R1 Distill Qwen 7B7.6BQ8_08.7 GBQ6_K15details
Falcon H1R 7B7.6BQ8_08.5 GBQ8_015details
Olmo 3 7B Instruct7.3BQ6_K8.7 GBQ3_K_M20details
Olmo 3 1025 7B7.3BQ6_K8.7 GBQ3_K_M20details
Olmo 3 7B Think7.3BQ6_K8.7 GBQ3_K_M20details
Mistral 7B v0.17.2BQ8_08.6 GBQ5_K_M16details
Mistral 7B Instruct v0.27.2BQ8_08.6 GBQ5_K_M16details
Mistral 7B Instruct v0.17.2BQ8_08.6 GBQ5_K_M16details
falcon 7b7.2BQ4_K_M9.6 GBno24details
falcon 7b instruct7.2BQ4_K_M9.4 GBno25details
granite 4.0 h tiny6.9B (1.5B act.)Q8_07.5 GBQ8_077details
OLMoE 1B 7B 09246.9B (1.3B act.)Q8_08.4 GBQ5_K_M88details
deepseek coder 7b instruct v1.56.9BQ6_K9.6 GBno21details
deepseek coder 6.7b instruct6.7BQ6_K9.6 GBno22details
deepseek coder 6.7b base6.7BQ6_K9.6 GBno22details
Llama 2 7b hf6.7BQ6_K9.6 GBno22details
Yi 6B 200K6.1BQ8_07.0 GBQ8_019details
Yi 6B6.1BQ8_07.0 GBQ8_019details
Molmo2 4B4.9BQ8_06.2 GBQ8_026details
WeVisOne VL 4B4.8BQ8_06.2 GBQ8_026details
Mage VL Cowork4.7BQ8_06.2 GBQ8_026details
Qianfan OCR4.7BQ8_06.2 GBQ8_026details
Qwen3.5 4B4.7BF169.0 GBQ8_014details
S5 nano4.7BF169.2 GBQ8_014details
LightOnOCR 3 4B4.5BF169.0 GBQ8_014details
SingVL S1 4B4.5BF169.0 GBQ8_014details
Intern Decision 4B4.5BF169.0 GBQ8_014details
Qwen3 VL 4B Instruct4.4BF169.3 GBQ8_015details
gemma 3 4b it4.3BF168.4 GBF1615details
medgemma 1.5 4b it4.3BF168.4 GBF1615details
medgemma 4b it4.3BF168.4 GBF1615details
gemma 3 4b pt4.3BF169.2 GBF1614details
Phi 3.5 vision instruct4.1BQ8_07.6 GBno27details
Qwen3 4B4.0BF169.3 GBQ8_015details
Qwen3 4B Instruct 25074.0BF169.3 GBQ8_015details
granite vision 4.1 4b4.0BF167.6 GBF1618details
Phi 4 mini instruct3.8BF168.9 GBQ8_016details
Phi 3 mini 4k instruct3.8BQ8_07.2 GBno30details
Phi 3.5 mini instruct3.8BQ8_07.2 GBno30details
Phi 3 mini 128k instruct3.8BQ8_07.2 GBno30details
Qwen2.5 VL 3B Instruct3.8BF166.7 GBF1619details
granite 4.2 3b3.7BF168.0 GBQ8_016details
granite 4.1 3b llmon preview3.7BF168.0 GBQ8_016details
Raon 1 nano3.5BF167.0 GBF1619details
granite 4.1 3b3.4BF167.6 GBF1618details
granite 4.0 micro3.4BF167.6 GBF1618details
DeepSeek OCR 23.4B (1.2B act.)F166.6 GBF1658details
deepseek vl2 tiny3.4B (1.2B act.)F167.4 GBF1651details
Unlimited OCR 23.3B (1.1B act.)F166.6 GBF1659details
Unlimited OCR3.3B (1.1B act.)F166.6 GBF1660details
DeepSeek OCR3.3B (1.1B act.)F166.6 GBF1660details
granite 3.1 3b a800m instruct3.3B (883M act.)F167.1 GBF1668details
Llama 3.2 3B Instruct3.2BQ8_04.7 GBQ8_035details
Llama 3.2 3B3.2BF167.5 GBQ8_019details
Hermes 3 Llama 3.2 3B3.2BF167.5 GBQ8_019details
d1 3B3.1BQ8_03.4 GBQ8_042details
LFM2.5 VL 3B3.1BQ8_03.4 GBQ8_042details
Qwen2.5 3B Instruct3.1BF166.7 GBF1619details
SmolLM3 3B3.1BF166.9 GBF1619details
SmolLM3 3B Base3.1BF166.9 GBF1620details
phi 22.8BF168.2 GBno22details
LFM2.5 2.6B2.7BQ8_03.4 GBQ8_042details
granite 3.3 2b instruct2.5BF165.8 GBF1624details
MiniCPM5 2B2.5BF165.7 GBF1624details
WeVisOne VL 2B2.4BF165.4 GBF1629details
Qwen3.5 2B2.3BF164.5 GBF1632details
SmolVLM Instruct2.2BF165.4 GBQ8_033details
Qwen3 1.7B2.0BF164.8 GBF1635details
Qwen3.8 27B DSpine b162.0BF164.8 GBF1630details
d1 2B1.9BF167.2 GBF1617details
LFM3 VL 2B1.9BF167.2 GBF1617details
DeepSeek R1 Distill Qwen 1.5B1.8BF164.2 GBF1634details
LFM3 1.7B1.7BF167.2 GBF1617details
LFM3 1.7B Base1.7BF164.0 GBF1635details
SmolLM2 1.7B Instruct1.7BF165.2 GBQ8_035details
SmolLM2 1.7B1.7BF165.2 GBQ8_035details
SmolLM 1.7B1.7BF165.2 GBQ8_035details
granite 4.0 1b1.6BF164.3 GBF1637details
Qwen2.5 1.5B Instruct1.5BF163.8 GBF1639details
OLMo 2 0425 1B1.5BF164.4 GBF1640details
deepseek coder 1.3b instruct1.3BF164.5 GBF1645details
Llama 3.2 1B1.2BF163.2 GBF1649details
Llama 3.2 1B Instruct1.2BF165.5 GBF1624details
LFM2.5 1.2B Instruct1.2BF165.0 GBF1626details
HunyuanOCR1.1BF163.5 GBF1654details
PaddleOCR VL 1.6959MF162.5 GBF1663details
Qwen3.5 0.8B873MF162.4 GBF1679details
LightOnOCR 3 0.8B853MF162.4 GBF1679details
FastVLM 0.5B759MF162.2 GBF1679details
Qwen3 0.6B752MF162.7 GBF16100details
GOT OCR2 0716MF162.8 GBF1684details
Nemotron 3.5 Parse631MF161.9 GBF1692details
GOT OCR 2.0 hf561MF162.5 GBF16107details
SmolVLM2 500M Video Instruct507MF161.6 GBF16146details
SmolVLM 500M Instruct507MF161.6 GBF16146details
Qwen2.5 0.5B Instruct494MF161.7 GBF16121details
LFM2.5 VL 450M449MF161.9 GBF1684details
SmolLM2 360M Instruct362MF161.5 GBF16165details
SmolLM2 360M362MF161.5 GBF16165details
LFM2.5 350M354MF161.9 GBF1684details
LFM2 350M354MF161.4 GBF16169details

Just too big at 4-bit

These need up to 35% more memory than the Mac M4 16GB has at Q4_K_M and 8K context. A smaller quant or shorter context may work.

ModelQ4_K_M needsShort by
Moonlight 16B A3B Instruct10.8 GB0.0 GB
Kimi VL A3B Thinking 250610.8 GB0.0 GB
Kimi VL A3B Instruct10.8 GB0.0 GB
glm 4 9b11.5 GB0.8 GB
gpt oss safeguard 20b11.8 GB1.1 GB
gpt oss 20b11.8 GB1.1 GB
deepseek moe 16b base12.5 GB1.7 GB
Groot 19B A3B 26102312.7 GB1.9 GB
ERNIE 4.5 21B A3B PT13.5 GB2.7 GB
ERNIE 4.5 21B A3B Thinking13.5 GB2.7 GB
LFM2.5 24B A2B14.3 GB3.6 GB

Questions

What is the biggest model that runs on the Mac M4 16GB?

Of the models we track, the largest that fits at 8K context is Trinity Mini (26B) at Q2_K, needing about 9.8 GB.

How much of the Mac M4 16GB's memory can a model use?

About 11 GB of 16 GB. macOS lets the GPU use about 67% of unified memory by default; this can be raised with the iogpu.wired_limit_mb setting.

Spec source: Apple specifications. Model data: Hugging Face Hub API. Speed is a bandwidth ceiling (bandwidth ÷ bytes read per token), not a benchmark. Methodology · Report an error.