What are the best models to run on Apple M4 (16GB)?

Based on model rankings and 12 GB usable VRAM: Top Overall: Phi-4 Reasoning Plus (INT4) ~8 tok/s. Top Coding: K2 Horizon 7B (INT4) ~23 tok/s. Optimal Daily Driver: Qwen3.5-9B (INT4) ~18 tok/s. Maximum Model Size: ERNIE-4.5-21B-A3B (INT4) ~36 tok/s.

How are VRAM, throughput, and latency calculated?

Calculations are determined using apxml.com modeling, combining LLM architecture analysis with GPU hardware physics. VRAM and Context Scaling: Memory requirements are determined from model architecture, including parameter count, layer configuration, and attention structure. The KV cache footprint scales deterministically based on context length, batch size, and cache precision. Throughput (TPS): Token generation is calculated in the memory-bandwidth-bound decode phase, measuring the time required to stream model weights and KV cache relative to hardware memory bandwidth. Latency (TTFT): Time to first token is calculated from the compute operations required during prompt prefill relative to hardware tensor compute capacity.

How accurate are these estimates?

Calculations are calibrated against empirical ground-truth benchmark runs across diverse architectures and hardware. Because the underlying mechanics are benchmark-calibrated, estimates remain accurate across different configurations without needing to run physical tests for every scenario. VRAM Accuracy: Memory estimates are deterministic and closely match actual allocation. Minor differences typically stay within 3% to 5% due to CUDA runtime overhead or memory allocator fragmentation. Speed and Latency Accuracy: Throughput and TTFT typically vary within 5% to 10% depending on the specific inference runtime, kernel implementations, and system power limits.

Best LLMs to Run on Apple M4 (16GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#102	Phi-4 Reasoning Plus	14B	INT4	8 tok/s	~39.4s
#121	Qwen3.5-9B	9B	INT4 Q8	18 tok/s 10 tok/s	~21.7s ~23.6s
#127	K2 Horizon 7B	7B	INT4 Q8	23 tok/s 13 tok/s	~16.9s ~17.0s
#130	Gemma 4 12B	11.95B	INT4	9 tok/s	~31.2s
#136	Qwen3.5-4B	4B	INT4 Q8 FP16	37 tok/s 22 tok/s 12 tok/s	~9.9s ~9.9s ~10.0s
#140	GLM-4V	9B	INT4 Q8	13 tok/s 8 tok/s	~21.7s ~25.5s
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	6 tok/s 3 tok/s 2 tok/s	~68.3s ~68.4s ~74.7s
#152	Granite 4.2 8B	8B	INT4 Q8	20 tok/s 12 tok/s	~19.3s ~19.3s
#157	Gemma 4 E4B	8B	INT4 Q8	20 tok/s 12 tok/s	~19.3s ~19.3s
#164	Phi-4	14B	INT4	12 tok/s	~33.5s
#176	DeepSeek-R1 14B	14B	INT4	13 tok/s	~33.5s
#177	Ministral 3 14B	14B	INT4	8 tok/s	~39.4s
#180	Gemma 3 12B	12B	INT4	14 tok/s	~28.8s
#185	Ministral 3 8B	8B	INT4 Q8	14 tok/s 9 tok/s	~19.3s ~21.0s
#187	OLMo 3 7B Think	7B	INT4 Q8	16 tok/s 10 tok/s	~17.0s ~18.5s
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	64 tok/s 41 tok/s 23 tok/s	~5.3s ~5.3s ~5.3s
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	64 tok/s 41 tok/s 23 tok/s	~5.3s ~5.3s ~5.3s
#190	Qwen2.5-14B	14B	INT4	12 tok/s	~33.5s
#193	Gemma 4 E2B	5.1B	INT4 Q8 FP16	30 tok/s 18 tok/s 9 tok/s	~12.5s ~12.5s ~13.7s
#195	Ministral 3 3B	3B	INT4 Q8 FP16	29 tok/s 21 tok/s 13 tok/s	~7.6s ~7.6s ~7.6s
#196	DeepSeek-R1 8B	8B	INT4 Q8	21 tok/s 12 tok/s	~19.3s ~19.3s
#198	Gemma 2 9B	9B	INT4 Q8	18 tok/s 10 tok/s	~21.7s ~23.6s
#201	Qwen3-4B	4B	INT4 Q8 FP16	37 tok/s 22 tok/s 12 tok/s	~9.9s ~9.9s ~10.0s
#202	Qwen3-14B	14B	INT4	12 tok/s	~33.5s
#203	Gemma 3 4B	4B	INT4 Q8 FP16	37 tok/s 22 tok/s 12 tok/s	~9.9s ~9.9s ~10.0s
#204	DeepSeek-R1 7B	7B	INT4 Q8	24 tok/s 14 tok/s	~16.9s ~17.0s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	120 tok/s 84 tok/s 51 tok/s	~2.1s ~2.1s ~2.1s
#207	Llama 3.1 8B	8B	INT4 Q8	20 tok/s 12 tok/s	~19.3s ~19.3s
#207	Qwen3-8B	8B	INT4 Q8	20 tok/s 12 tok/s	~19.3s ~19.3s
#212	Llama 3 8B	8B	INT4 Q8	20 tok/s 12 tok/s	~19.3s ~19.3s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	72 tok/s 47 tok/s 26 tok/s	~4.5s ~4.5s ~4.5s
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	161 tok/s 137 tok/s 101 tok/s	~693 ms ~694 ms ~697 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	143 tok/s 104 tok/s 64 tok/s	~1.6s ~1.6s ~1.6s
#217	Phi-4-Mini	3.8B	INT4 Q8 FP16	39 tok/s 24 tok/s 13 tok/s	~9.5s ~9.5s ~9.5s
#218	Qwen2.5-7B	7B	INT4 Q8	23 tok/s 13 tok/s	~16.9s ~17.0s
#220	Ministral-8B-2410	8B	INT4 Q8	20 tok/s 12 tok/s	~19.3s ~19.3s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	79 tok/s 52 tok/s 29 tok/s	~4.0s ~4.0s ~4.0s
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
#223	Phi-3-mini	3.8B	INT4 Q8 FP16	39 tok/s 24 tok/s 13 tok/s	~9.5s ~9.5s ~9.5s
#224	Gemma 3n E2B IT	6B (2B active)	INT4 Q8	29 tok/s 24 tok/s	~5.6s ~5.6s
#225	Gemma 2 2B	2B	INT4 Q8 FP16	64 tok/s 41 tok/s 23 tok/s	~5.3s ~5.3s ~5.3s
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8	22 tok/s 13 tok/s	~17.6s ~17.7s
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	103 tok/s 71 tok/s 42 tok/s	~2.7s ~2.7s ~2.7s
#228	Phi-3-medium	14B	INT4	12 tok/s	~33.5s
#230	OLMo 3 7B Instruct	7B	INT4 Q8	16 tok/s 10 tok/s	~17.0s ~18.5s
#231	Qwen2-7B	7B	INT4 Q8	23 tok/s 13 tok/s	~16.9s ~17.0s
#232	Phi-3-small	7B	INT4 Q8	23 tok/s 13 tok/s	~16.9s ~17.0s
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8	22 tok/s 13 tok/s	~17.6s ~17.7s
#234	Gemma 3 1B	1B	INT4 Q8 FP16	103 tok/s 71 tok/s 42 tok/s	~2.7s ~2.7s ~2.7s
#236	Gemma 1 7B	7B	INT4 Q8	16 tok/s 10 tok/s	~17.0s ~18.5s
#237	Gemma 1 2B	2B	INT4 Q8 FP16	70 tok/s 43 tok/s 23 tok/s	~5.3s ~5.3s ~5.3s
#238	ChatGLM3-6B	6B	INT4 Q8	18 tok/s 12 tok/s	~14.6s ~14.6s
#239	ChatGLM2-6B	6B	INT4 Q8	18 tok/s 12 tok/s	~14.6s ~14.6s
#240	ChatGLM-6B	6B	INT4 Q8	18 tok/s 12 tok/s	~14.6s ~14.6s
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	196 tok/s 155 tok/s 107 tok/s	~772 ms ~773 ms ~776 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	157 tok/s 117 tok/s 74 tok/s	~1.3s ~1.3s ~1.3s
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	157 tok/s 117 tok/s 74 tok/s	~1.3s ~1.3s ~1.3s
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	152 tok/s 128 tok/s 93 tok/s	~773 ms ~775 ms ~778 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	152 tok/s 128 tok/s 93 tok/s	~773 ms ~775 ms ~778 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	143 tok/s 104 tok/s 64 tok/s	~1.6s ~1.6s ~1.6s
-	Falcon-1B	1B	INT4 Q8 FP16	113 tok/s 75 tok/s 43 tok/s	~2.7s ~2.7s ~2.7s
-	Falcon3-1B	1B	INT4 Q8 FP16	103 tok/s 71 tok/s 42 tok/s	~2.7s ~2.7s ~2.7s
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	103 tok/s 71 tok/s 42 tok/s	~2.7s ~2.7s ~2.7s
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	79 tok/s 52 tok/s 29 tok/s	~4.0s ~4.0s ~4.0s
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	79 tok/s 52 tok/s 29 tok/s	~4.0s ~4.0s ~4.0s
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	72 tok/s 47 tok/s 26 tok/s	~4.5s ~4.5s ~4.5s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	69 tok/s 45 tok/s 25 tok/s	~4.7s ~4.7s ~4.8s
-	Falcon-3B	3B	INT4 Q8 FP16	51 tok/s 30 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	49 tok/s 40 tok/s 26 tok/s	~3.5s ~3.5s ~3.5s
-	Phi-1	1.3B	INT4 Q8 FP16	49 tok/s 40 tok/s 26 tok/s	~3.5s ~3.5s ~3.5s
-	Phi-1.5	1.3B	INT4 Q8 FP16	49 tok/s 40 tok/s 26 tok/s	~3.5s ~3.5s ~3.5s
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	48 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	48 tok/s	~7.4s
-	Falcon3-3B	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	Liquid D1 3B	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	47 tok/s 29 tok/s 16 tok/s	~7.6s ~7.6s ~7.6s
-	FrogNano 4B	4B	INT4 Q8 FP16	37 tok/s 22 tok/s 12 tok/s	~9.9s ~9.9s ~10.0s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	37 tok/s 22 tok/s 12 tok/s	~9.9s ~9.9s ~10.0s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	36 tok/s	~10.9s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	36 tok/s	~10.9s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	35 tok/s 28 tok/s 18 tok/s	~5.3s ~5.3s ~5.3s
-	Phi-2	2.7B	INT4 Q8 FP16	30 tok/s 23 tok/s 14 tok/s	~6.9s ~6.9s ~6.9s
-	MaLLaM-3B	3B	INT4 Q8 FP16	29 tok/s 21 tok/s 13 tok/s	~7.6s ~7.6s ~7.6s
-	SmolLM3 3B	3B	INT4 Q8 FP16	29 tok/s 21 tok/s 13 tok/s	~7.6s ~7.6s ~7.6s
-	Falcon-7B	7B	INT4 Q8	24 tok/s 14 tok/s	~16.9s ~17.0s
-	Falcon3-7B	7B	INT4 Q8	23 tok/s 13 tok/s	~16.9s ~17.0s
-	Hy-MT2 7B	7B	INT4 Q8	23 tok/s 13 tok/s	~16.9s ~17.0s
-	Mistral-7B-v0.1	7.3B	INT4 Q8	22 tok/s 13 tok/s	~17.6s ~17.7s
-	Youtu Parsing Omni	8B	INT4 Q8	21 tok/s 12 tok/s	~19.3s ~19.3s
-	ChatGLM3-6B-32K	6B	INT4 Q8	18 tok/s 12 tok/s	~14.6s ~14.6s
-	LensVLM 9B	9B	INT4 Q8	18 tok/s 10 tok/s	~21.7s ~23.6s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	18 tok/s 10 tok/s	~21.7s ~23.6s
-	Yi-6B	6B	INT4 Q8	18 tok/s 12 tok/s	~14.6s ~14.6s
-	Falcon3-10B	10B	INT4 Q8	17 tok/s 9 tok/s	~24.0s ~28.3s
-	Falcon2-11B	11B	INT4 Q8	16 tok/s 8 tok/s	~26.4s ~31.1s
-	Hunyuan Lite	7B	INT4 Q8	16 tok/s 10 tok/s	~17.0s ~18.5s
-	OLMo 3 7B Base	7B	INT4 Q8	16 tok/s 10 tok/s	~17.0s ~18.5s
-	SEA-LION-7B	7.1B	INT4 Q8	16 tok/s 10 tok/s	~17.2s ~18.7s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	16 tok/s 10 tok/s	~17.2s ~18.7s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	14 tok/s 9 tok/s	~19.3s ~21.0s
-	Typhoon-2-8B	8B	INT4 Q8	14 tok/s 9 tok/s	~19.3s ~21.0s
-	GLM-4-9B	9B	INT4 Q8	13 tok/s 8 tok/s	~21.7s ~25.5s
-	GLM-4-9B-Chat	9B	INT4 Q8	13 tok/s 8 tok/s	~21.7s ~25.5s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	13 tok/s 8 tok/s	~21.7s ~25.5s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4	13 tok/s	~22.1s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4	13 tok/s	~22.1s
-	Yi-9B	9B	INT4 Q8	13 tok/s 8 tok/s	~21.7s ~25.5s
-	GreenMind-14B-R1	14B	INT4	8 tok/s	~39.4s

Apple M4 (16GB)

VRAM Scaling by Context Length

Best Models to Run on Apple M4 (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions