What are the best models to run on Apple M2 Pro (16GB)?

Based on model rankings and 12 GB usable VRAM: Top Overall: Phi-4 Reasoning Plus (INT4) ~13 tok/s. Top Coding: K2 Horizon 7B (INT4) ~37 tok/s. Optimal Daily Driver: Gemma 4 12B (INT4) ~15 tok/s. Maximum Model Size: ERNIE-4.5-21B-A3B (INT4) ~55 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 Models to Run on Apple M2 Pro (16GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#95	Phi-4 Reasoning Plus	14B	INT4	13 tok/s	~29.5s
#108	Gemma 4 12B	11.95B	INT4	15 tok/s	~23.3s
#119	K2 Horizon 7B	7B	INT4 Q8	37 tok/s 22 tok/s	~12.7s ~12.7s
#129	Qwen3.5-4B	4B	INT4 Q8 FP16	57 tok/s 36 tok/s 19 tok/s	~7.4s ~7.4s ~7.5s
#130	Qwen3.5-9B	9B	INT4 Q8	29 tok/s 16 tok/s	~16.2s ~17.6s
#139	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	9 tok/s 5 tok/s 3 tok/s	~51.1s ~51.2s ~55.8s
#147	Granite 4.2 8B	8B	INT4 Q8	33 tok/s 19 tok/s	~14.4s ~14.4s
#154	Gemma 4 E4B	8B	INT4 Q8	33 tok/s 19 tok/s	~14.4s ~14.4s
#158	GLM-4V	9B	INT4 Q8	21 tok/s 13 tok/s	~16.2s ~19.1s
#160	Phi-4	14B	INT4	20 tok/s	~25.0s
#175	Ministral 3 14B	14B	INT4	13 tok/s	~29.5s
#176	DeepSeek-R1 14B	14B	INT4	20 tok/s	~25.0s
#179	Gemma 3 12B	12B	INT4	23 tok/s	~21.5s
#183	Ministral 3 8B	8B	INT4 Q8	23 tok/s 15 tok/s	~14.4s ~15.7s
#185	OLMo 3 7B Think	7B	INT4 Q8	25 tok/s 17 tok/s	~12.7s ~13.8s
#186	Qwen3.5-2B	2B	INT4 Q8 FP16	93 tok/s 62 tok/s 36 tok/s	~3.9s ~3.9s ~4.0s
#187	MiniCPM5-2B	2B	INT4 Q8 FP16	93 tok/s 62 tok/s 36 tok/s	~3.9s ~3.9s ~4.0s
#188	Qwen2.5-14B	14B	INT4	20 tok/s	~25.0s
#190	Gemma 4 E2B	5.1B	INT4 Q8 FP16	47 tok/s 29 tok/s 15 tok/s	~9.3s ~9.4s ~10.2s
#193	Ministral 3 3B	3B	INT4 Q8 FP16	45 tok/s 34 tok/s 21 tok/s	~5.7s ~5.7s ~5.7s
#194	DeepSeek-R1 8B	8B	INT4 Q8	33 tok/s 20 tok/s	~14.4s ~14.4s
#195	Gemma 2 9B	9B	INT4 Q8	29 tok/s 16 tok/s	~16.2s ~17.6s
#198	Qwen3-4B	4B	INT4 Q8 FP16	57 tok/s 36 tok/s 19 tok/s	~7.4s ~7.4s ~7.5s
#199	Qwen3-14B	14B	INT4	20 tok/s	~25.0s
#200	Gemma 3 4B	4B	INT4 Q8 FP16	57 tok/s 36 tok/s 19 tok/s	~7.4s ~7.4s ~7.5s
#201	Qwen2.5-7B	7B	INT4 Q8	37 tok/s 22 tok/s	~12.7s ~12.7s
#202	Qwen3-8B	8B	INT4 Q8	33 tok/s 19 tok/s	~14.4s ~14.4s
#203	DeepSeek-R1 7B	7B	INT4 Q8	38 tok/s 22 tok/s	~12.7s ~12.7s
#204	Llama 3.1 8B	8B	INT4 Q8	33 tok/s 19 tok/s	~14.4s ~14.4s
#207	Llama 3 8B	8B	INT4 Q8	33 tok/s 19 tok/s	~14.4s ~14.4s
#209	Qwen3-1.7B	1.7B	INT4 Q8 FP16	103 tok/s 71 tok/s 41 tok/s	~3.4s ~3.4s ~3.4s
#210	Gemma 3 270M	0.27B	INT4 Q8 FP16	199 tok/s 176 tok/s 137 tok/s	~520 ms ~521 ms ~523 ms
#211	Qwen3-0.6B	0.6B	INT4 Q8 FP16	181 tok/s 141 tok/s 94 tok/s	~1.2s ~1.2s ~1.2s
#212	Phi-4-Mini	3.8B	INT4 Q8 FP16	60 tok/s 37 tok/s 20 tok/s	~7.1s ~7.1s ~7.1s
#213	Ministral-8B-2410	8B	INT4 Q8	33 tok/s 19 tok/s	~14.4s ~14.4s
#215	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	112 tok/s 78 tok/s 46 tok/s	~3.0s ~3.0s ~3.0s
#216	Llama 3.2 3B	3B	INT4 Q8 FP16	71 tok/s 45 tok/s 25 tok/s	~5.7s ~5.7s ~5.7s
#217	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	158 tok/s 119 tok/s 76 tok/s	~1.6s ~1.6s ~1.6s
#218	Phi-3-mini	3.8B	INT4 Q8 FP16	60 tok/s 37 tok/s 20 tok/s	~7.1s ~7.1s ~7.1s
#219	Gemma 3n E2B IT	6B (2B active)	INT4 Q8	45 tok/s 37 tok/s	~4.2s ~4.2s
#220	Gemma 2 2B	2B	INT4 Q8 FP16	93 tok/s 62 tok/s 36 tok/s	~3.9s ~3.9s ~4.0s
#221	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8	35 tok/s 21 tok/s	~13.2s ~13.2s
#222	Phi-3-medium	14B	INT4	20 tok/s	~25.0s
#223	Llama 3.2 1B	1B	INT4 Q8 FP16	140 tok/s 102 tok/s 64 tok/s	~2.0s ~2.0s ~2.0s
#225	OLMo 3 7B Instruct	7B	INT4 Q8	25 tok/s 17 tok/s	~12.7s ~13.8s
#226	Qwen2-7B	7B	INT4 Q8	37 tok/s 22 tok/s	~12.7s ~12.7s
#227	Phi-3-small	7B	INT4 Q8	37 tok/s 22 tok/s	~12.7s ~12.7s
#228	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8	35 tok/s 21 tok/s	~13.2s ~13.2s
#230	Gemma 3 1B	1B	INT4 Q8 FP16	140 tok/s 102 tok/s 64 tok/s	~2.0s ~2.0s ~2.0s
#231	Gemma 1 7B	7B	INT4 Q8	25 tok/s 17 tok/s	~12.7s ~13.8s
#232	Gemma 1 2B	2B	INT4 Q8 FP16	100 tok/s 65 tok/s 37 tok/s	~3.9s ~3.9s ~4.0s
#233	ChatGLM3-6B	6B	INT4 Q8	28 tok/s 20 tok/s	~10.9s ~10.9s
#234	ChatGLM2-6B	6B	INT4 Q8	28 tok/s 20 tok/s	~10.9s ~10.9s
#235	ChatGLM-6B	6B	INT4 Q8	28 tok/s 20 tok/s	~10.9s ~10.9s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	195 tok/s 155 tok/s 106 tok/s	~972 ms ~973 ms ~976 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	195 tok/s 155 tok/s 106 tok/s	~972 ms ~973 ms ~976 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	189 tok/s 167 tok/s 129 tok/s	~580 ms ~581 ms ~583 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	189 tok/s 167 tok/s 129 tok/s	~580 ms ~581 ms ~583 ms
-	Falcon-1B	1B	INT4 Q8 FP16	150 tok/s 107 tok/s 66 tok/s	~2.0s ~2.0s ~2.0s
-	Falcon3-1B	1B	INT4 Q8 FP16	140 tok/s 102 tok/s 64 tok/s	~2.0s ~2.0s ~2.0s
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	112 tok/s 78 tok/s 46 tok/s	~3.0s ~3.0s ~3.0s
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	112 tok/s 78 tok/s 46 tok/s	~3.0s ~3.0s ~3.0s
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	103 tok/s 71 tok/s 41 tok/s	~3.4s ~3.4s ~3.4s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	100 tok/s 68 tok/s 39 tok/s	~3.5s ~3.5s ~3.6s
-	Falcon-3B	3B	INT4 Q8 FP16	76 tok/s 47 tok/s 26 tok/s	~5.7s ~5.7s ~5.7s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	74 tok/s 61 tok/s 41 tok/s	~2.6s ~2.6s ~2.6s
-	Phi-1	1.3B	INT4 Q8 FP16	74 tok/s 61 tok/s 41 tok/s	~2.6s ~2.6s ~2.6s
-	Phi-1.5	1.3B	INT4 Q8 FP16	74 tok/s 61 tok/s 41 tok/s	~2.6s ~2.6s ~2.6s
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	72 tok/s 46 tok/s 25 tok/s	~5.7s ~5.7s ~5.7s
-	Falcon3-3B	3B	INT4 Q8 FP16	71 tok/s 45 tok/s 25 tok/s	~5.7s ~5.7s ~5.7s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	71 tok/s 45 tok/s 25 tok/s	~5.7s ~5.7s ~5.7s
-	FrogNano 4B	4B	INT4 Q8 FP16	57 tok/s 36 tok/s 19 tok/s	~7.4s ~7.4s ~7.5s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	57 tok/s 36 tok/s 19 tok/s	~7.4s ~7.4s ~7.5s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	55 tok/s 44 tok/s 29 tok/s	~3.9s ~3.9s ~4.0s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	55 tok/s	~8.2s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	55 tok/s	~8.2s
-	Phi-2	2.7B	INT4 Q8 FP16	47 tok/s 36 tok/s 23 tok/s	~5.2s ~5.2s ~5.2s
-	MaLLaM-3B	3B	INT4 Q8 FP16	45 tok/s 34 tok/s 21 tok/s	~5.7s ~5.7s ~5.7s
-	SmolLM3 3B	3B	INT4 Q8 FP16	45 tok/s 34 tok/s 21 tok/s	~5.7s ~5.7s ~5.7s
-	Falcon-7B	7B	INT4 Q8	38 tok/s 22 tok/s	~12.7s ~12.7s
-	Falcon3-7B	7B	INT4 Q8	37 tok/s 22 tok/s	~12.7s ~12.7s
-	Hy-MT2 7B	7B	INT4 Q8	37 tok/s 22 tok/s	~12.7s ~12.7s
-	Mistral-7B-v0.1	7.3B	INT4 Q8	35 tok/s 21 tok/s	~13.2s ~13.2s
-	LensVLM 9B	9B	INT4 Q8	29 tok/s 16 tok/s	~16.2s ~17.6s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	29 tok/s 16 tok/s	~16.2s ~17.6s
-	ChatGLM3-6B-32K	6B	INT4 Q8	28 tok/s 20 tok/s	~10.9s ~10.9s
-	Yi-6B	6B	INT4 Q8	28 tok/s 20 tok/s	~10.9s ~10.9s
-	Falcon3-10B	10B	INT4 Q8	27 tok/s 14 tok/s	~18.0s ~21.2s
-	Falcon2-11B	11B	INT4 Q8	26 tok/s 13 tok/s	~19.7s ~23.2s
-	Hunyuan Lite	7B	INT4 Q8	25 tok/s 17 tok/s	~12.7s ~13.8s
-	OLMo 3 7B Base	7B	INT4 Q8	25 tok/s 17 tok/s	~12.7s ~13.8s
-	SEA-LION-7B	7.1B	INT4 Q8	25 tok/s 16 tok/s	~12.9s ~14.0s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	25 tok/s 16 tok/s	~12.9s ~14.0s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	23 tok/s 15 tok/s	~14.4s ~15.7s
-	Typhoon-2-8B	8B	INT4 Q8	23 tok/s 15 tok/s	~14.4s ~15.7s
-	GLM-4-9B	9B	INT4 Q8	21 tok/s 13 tok/s	~16.2s ~19.1s
-	GLM-4-9B-Chat	9B	INT4 Q8	21 tok/s 13 tok/s	~16.2s ~19.1s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	21 tok/s 13 tok/s	~16.2s ~19.1s
-	Yi-9B	9B	INT4 Q8	21 tok/s 13 tok/s	~16.2s ~19.1s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4	20 tok/s	~16.6s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4	20 tok/s	~16.6s
-	GreenMind-14B-R1	14B	INT4	13 tok/s	~29.5s

Apple M2 Pro (16GB)

VRAM Scaling by Context Length

Best Models to Run on Apple M2 Pro (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions