What are the best models to run on Apple M5 (16GB)?

Based on model rankings and 12 GB usable VRAM: Top Overall: Phi-4 Reasoning Plus (INT4) ~10 tok/s. Top Coding: K2 Horizon 7B (INT4) ~29 tok/s. Optimal Daily Driver: Gemma 4 12B (INT4) ~12 tok/s. Maximum Model Size: ERNIE-4.5-21B-A3B (INT4) ~44 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 M5 (16GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#98	Phi-4 Reasoning Plus	14B	INT4	10 tok/s	~40.3s
#109	Gemma 4 12B	11.95B	INT4	12 tok/s	~31.9s
#113	Qwen3.5-9B	9B	INT4 Q8	23 tok/s 13 tok/s	~22.2s ~24.1s
#119	K2 Horizon 7B	7B	INT4 Q8	29 tok/s 17 tok/s	~17.3s ~17.4s
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	7 tok/s 4 tok/s 2 tok/s	~69.9s ~70.0s ~76.3s
#133	Qwen3.5-4B	4B	INT4 Q8 FP16	46 tok/s 28 tok/s 15 tok/s	~10.2s ~10.2s ~10.2s
#146	Granite 4.2 8B	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#152	Gemma 4 E4B	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#158	GLM-4V	9B	INT4 Q8	16 tok/s 10 tok/s	~22.2s ~26.1s
#160	Phi-4	14B	INT4	15 tok/s	~34.2s
#167	OLMo 3 7B Think	7B	INT4 Q8	20 tok/s 13 tok/s	~17.4s ~18.9s
#174	Ministral 3 14B	14B	INT4	10 tok/s	~40.3s
#175	DeepSeek-R1 14B	14B	INT4	16 tok/s	~34.2s
#179	Gemma 3 12B	12B	INT4	18 tok/s	~29.4s
#183	Ministral 3 8B	8B	INT4 Q8	18 tok/s 12 tok/s	~19.8s ~21.5s
#185	Qwen3.5-2B	2B	INT4 Q8 FP16	77 tok/s 50 tok/s 28 tok/s	~5.4s ~5.4s ~5.4s
#186	MiniCPM5-2B	2B	INT4 Q8 FP16	77 tok/s 50 tok/s 28 tok/s	~5.4s ~5.4s ~5.4s
#187	Qwen2.5-14B	14B	INT4	15 tok/s	~34.2s
#189	Gemma 4 E2B	5.1B	INT4 Q8 FP16	38 tok/s 23 tok/s 11 tok/s	~12.8s ~12.8s ~14.0s
#190	Qwen3-4B	4B	INT4 Q8 FP16	46 tok/s 28 tok/s 15 tok/s	~10.2s ~10.2s ~10.2s
#192	Ministral 3 3B	3B	INT4 Q8 FP16	35 tok/s 27 tok/s 16 tok/s	~7.8s ~7.8s ~7.8s
#193	DeepSeek-R1 8B	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#195	Qwen3-14B	14B	INT4	15 tok/s	~34.2s
#196	Gemma 2 9B	9B	INT4 Q8	23 tok/s 13 tok/s	~22.2s ~24.1s
#199	Gemma 3 4B	4B	INT4 Q8 FP16	46 tok/s 28 tok/s 15 tok/s	~10.2s ~10.2s ~10.2s
#200	Qwen2.5-7B	7B	INT4 Q8	29 tok/s 17 tok/s	~17.3s ~17.4s
#201	Qwen3-8B	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#202	DeepSeek-R1 7B	7B	INT4 Q8	30 tok/s 17 tok/s	~17.3s ~17.4s
#203	Llama 3.1 8B	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#207	Llama 3 8B	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#208	Qwen3-1.7B	1.7B	INT4 Q8 FP16	86 tok/s 57 tok/s 33 tok/s	~4.6s ~4.6s ~4.6s
#209	Gemma 3 270M	0.27B	INT4 Q8 FP16	180 tok/s 156 tok/s 118 tok/s	~709 ms ~710 ms ~712 ms
#210	Qwen3-0.6B	0.6B	INT4 Q8 FP16	161 tok/s 121 tok/s 78 tok/s	~1.6s ~1.6s ~1.6s
#211	Phi-4-Mini	3.8B	INT4 Q8 FP16	48 tok/s 29 tok/s 16 tok/s	~9.7s ~9.7s ~9.7s
#213	Ministral-8B-2410	8B	INT4 Q8	26 tok/s 15 tok/s	~19.7s ~19.8s
#214	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	94 tok/s 63 tok/s 37 tok/s	~4.1s ~4.1s ~4.1s
#215	Llama 3.2 3B	3B	INT4 Q8 FP16	58 tok/s 36 tok/s 20 tok/s	~7.8s ~7.8s ~7.8s
#216	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	138 tok/s 100 tok/s 62 tok/s	~2.2s ~2.2s ~2.2s
#217	Phi-3-mini	3.8B	INT4 Q8 FP16	48 tok/s 29 tok/s 16 tok/s	~9.7s ~9.7s ~9.7s
#218	Gemma 3n E2B IT	6B (2B active)	INT4 Q8	35 tok/s 29 tok/s	~5.8s ~5.8s
#219	Gemma 2 2B	2B	INT4 Q8 FP16	77 tok/s 50 tok/s 28 tok/s	~5.4s ~5.4s ~5.4s
#220	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8	28 tok/s 16 tok/s	~18.0s ~18.1s
#221	Llama 3.2 1B	1B	INT4 Q8 FP16	120 tok/s 85 tok/s 51 tok/s	~2.8s ~2.8s ~2.8s
#222	Phi-3-medium	14B	INT4	15 tok/s	~34.2s
#224	OLMo 3 7B Instruct	7B	INT4 Q8	20 tok/s 13 tok/s	~17.4s ~18.9s
#225	Qwen2-7B	7B	INT4 Q8	29 tok/s 17 tok/s	~17.3s ~17.4s
#226	Phi-3-small	7B	INT4 Q8	29 tok/s 17 tok/s	~17.3s ~17.4s
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8	28 tok/s 16 tok/s	~18.0s ~18.1s
#229	Gemma 3 1B	1B	INT4 Q8 FP16	120 tok/s 85 tok/s 51 tok/s	~2.8s ~2.8s ~2.8s
#230	Gemma 1 7B	7B	INT4 Q8	20 tok/s 13 tok/s	~17.4s ~18.9s
#231	Gemma 1 2B	2B	INT4 Q8 FP16	83 tok/s 53 tok/s 29 tok/s	~5.4s ~5.4s ~5.4s
#232	ChatGLM3-6B	6B	INT4 Q8	22 tok/s 15 tok/s	~15.0s ~15.0s
#233	ChatGLM2-6B	6B	INT4 Q8	22 tok/s 15 tok/s	~15.0s ~15.0s
#234	ChatGLM-6B	6B	INT4 Q8	22 tok/s 15 tok/s	~15.0s ~15.0s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	176 tok/s 135 tok/s 89 tok/s	~1.3s ~1.3s ~1.3s
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	176 tok/s 135 tok/s 89 tok/s	~1.3s ~1.3s ~1.3s
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	170 tok/s 146 tok/s 109 tok/s	~791 ms ~792 ms ~794 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	170 tok/s 146 tok/s 109 tok/s	~791 ms ~792 ms ~794 ms
-	Falcon-1B	1B	INT4 Q8 FP16	130 tok/s 89 tok/s 53 tok/s	~2.8s ~2.8s ~2.8s
-	Falcon3-1B	1B	INT4 Q8 FP16	120 tok/s 85 tok/s 51 tok/s	~2.8s ~2.8s ~2.8s
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	94 tok/s 63 tok/s 37 tok/s	~4.1s ~4.1s ~4.1s
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	94 tok/s 63 tok/s 37 tok/s	~4.1s ~4.1s ~4.1s
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	86 tok/s 57 tok/s 33 tok/s	~4.6s ~4.6s ~4.6s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	83 tok/s 55 tok/s 31 tok/s	~4.8s ~4.8s ~4.9s
-	Falcon-3B	3B	INT4 Q8 FP16	62 tok/s 37 tok/s 20 tok/s	~7.8s ~7.8s ~7.8s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	60 tok/s 49 tok/s 33 tok/s	~3.5s ~3.5s ~3.6s
-	Phi-1	1.3B	INT4 Q8 FP16	60 tok/s 49 tok/s 33 tok/s	~3.5s ~3.5s ~3.6s
-	Phi-1.5	1.3B	INT4 Q8 FP16	60 tok/s 49 tok/s 33 tok/s	~3.5s ~3.5s ~3.6s
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	58 tok/s 36 tok/s 20 tok/s	~7.8s ~7.8s ~7.8s
-	Falcon3-3B	3B	INT4 Q8 FP16	58 tok/s 36 tok/s 20 tok/s	~7.8s ~7.8s ~7.8s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	58 tok/s 36 tok/s 20 tok/s	~7.8s ~7.8s ~7.8s
-	FrogNano 4B	4B	INT4 Q8 FP16	46 tok/s 28 tok/s 15 tok/s	~10.2s ~10.2s ~10.2s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	46 tok/s 28 tok/s 15 tok/s	~10.2s ~10.2s ~10.2s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	44 tok/s	~11.2s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	44 tok/s	~11.2s
-	Phi-2	2.7B	INT4 Q8 FP16	38 tok/s 28 tok/s 18 tok/s	~7.1s ~7.1s ~7.1s
-	MaLLaM-3B	3B	INT4 Q8 FP16	35 tok/s 27 tok/s 16 tok/s	~7.8s ~7.8s ~7.8s
-	SmolLM3 3B	3B	INT4 Q8 FP16	35 tok/s 27 tok/s 16 tok/s	~7.8s ~7.8s ~7.8s
-	Falcon-7B	7B	INT4 Q8	30 tok/s 17 tok/s	~17.3s ~17.4s
-	Falcon3-7B	7B	INT4 Q8	29 tok/s 17 tok/s	~17.3s ~17.4s
-	Hy-MT2 7B	7B	INT4 Q8	29 tok/s 17 tok/s	~17.3s ~17.4s
-	Mistral-7B-v0.1	7.3B	INT4 Q8	28 tok/s 16 tok/s	~18.0s ~18.1s
-	LensVLM 9B	9B	INT4 Q8	23 tok/s 13 tok/s	~22.2s ~24.1s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	23 tok/s 13 tok/s	~22.2s ~24.1s
-	ChatGLM3-6B-32K	6B	INT4 Q8	22 tok/s 15 tok/s	~15.0s ~15.0s
-	Yi-6B	6B	INT4 Q8	22 tok/s 15 tok/s	~15.0s ~15.0s
-	Falcon3-10B	10B	INT4 Q8	21 tok/s 11 tok/s	~24.6s ~28.9s
-	Falcon2-11B	11B	INT4 Q8	20 tok/s 10 tok/s	~27.0s ~31.8s
-	Hunyuan Lite	7B	INT4 Q8	20 tok/s 13 tok/s	~17.4s ~18.9s
-	OLMo 3 7B Base	7B	INT4 Q8	20 tok/s 13 tok/s	~17.4s ~18.9s
-	SEA-LION-7B	7.1B	INT4 Q8	20 tok/s 13 tok/s	~17.6s ~19.1s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	20 tok/s 13 tok/s	~17.6s ~19.1s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	18 tok/s 12 tok/s	~19.8s ~21.5s
-	Typhoon-2-8B	8B	INT4 Q8	18 tok/s 12 tok/s	~19.8s ~21.5s
-	GLM-4-9B	9B	INT4 Q8	16 tok/s 10 tok/s	~22.2s ~26.1s
-	GLM-4-9B-Chat	9B	INT4 Q8	16 tok/s 10 tok/s	~22.2s ~26.1s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	16 tok/s 10 tok/s	~22.2s ~26.1s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4	16 tok/s	~22.7s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4	16 tok/s	~22.7s
-	Yi-9B	9B	INT4 Q8	16 tok/s 10 tok/s	~22.2s ~26.1s
-	GreenMind-14B-R1	14B	INT4	10 tok/s	~40.3s

Apple M5 (16GB)

VRAM Scaling by Context Length

Best Models to Run on Apple M5 (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions