What are the best models to run on Apple M2 Ultra (64GB)?

Based on model rankings and 48 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~39 tok/s. Top Coding: Qwen3.5-27B (INT4) ~39 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~50 tok/s. Maximum Model Size: AliceAI Foundation 80B A3B Base (INT4) ~117 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 Ultra (64GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4 Q8	39 tok/s 23 tok/s	~10.0s ~10.0s
#48	Agnes 3.0 Flash	33B	INT4 Q8	32 tok/s 19 tok/s	~12.1s ~12.1s
#72	K2 Horizon MoVA 36B A4B	36B (4B active)	INT4 Q8	129 tok/s 97 tok/s	~2.0s ~2.0s
#83	Qwen3.5-27B	27B	INT4 Q8	39 tok/s 23 tok/s	~10.0s ~10.0s
#92	Sarvam-30B	32B (2.4B active)	INT4 Q8	157 tok/s 127 tok/s	~1.4s ~1.4s
#94	Qwen3.6 35B A3B	35B (3B active)	INT4 Q8	145 tok/s 113 tok/s	~1.6s ~1.6s
#96	Muse Glimmer 30B	30B	INT4 Q8	35 tok/s 21 tok/s	~11.0s ~11.1s
#100	Phi-4 Reasoning Plus	14B	INT4 Q8 FP16	50 tok/s 34 tok/s 20 tok/s	~5.2s ~5.3s ~5.3s
#101	MiMo V2 Flash	15B (309B active)	INT4 Q8 FP16	4 tok/s 2 tok/s 1 tok/s	~108.9s ~109.2s ~109.6s
#108	Gemma 4 26B A4B	25.2B (3.8B active)	INT4 Q8	136 tok/s 102 tok/s	~1.8s ~1.8s
#109	Gemma 4 12B	11.95B	INT4 Q8 FP16	56 tok/s 39 tok/s 23 tok/s	~4.5s ~4.5s ~4.5s
#113	Qwen3.5-9B	9B	INT4 Q8 FP16	91 tok/s 59 tok/s 33 tok/s	~3.4s ~3.4s ~3.4s
#114	Qwen3.5-35B-A3B	35B (3B active)	INT4 Q8	145 tok/s 113 tok/s	~1.6s ~1.6s
#118	K2 Horizon 7B	7B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	34 tok/s 20 tok/s 11 tok/s	~10.7s ~10.7s ~10.8s
#129	Gemma 4 31B	30.7B	INT4 Q8	35 tok/s 20 tok/s	~11.3s ~11.3s
#131	Nemotron 3.5 Lightning	30B (3B active)	INT4 Q8	68 tok/s 62 tok/s	~1.6s ~1.6s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	147 tok/s 105 tok/s 65 tok/s	~1.6s ~1.6s ~1.6s
#133	Qwen3-30B-A3B	30B (3B active)	INT4 Q8	147 tok/s 114 tok/s	~1.6s ~1.6s
#136	GPT-OSS 20B	21B (3.6B active)	INT4 Q8 FP16	75 tok/s 66 tok/s 44 tok/s	~1.7s ~1.7s ~2.0s
#146	Granite 4.2 8B	8B	INT4 Q8 FP16	98 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#150	Gemma 4 E4B	8B	INT4 Q8 FP16	98 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#151	Nemotron 3.5 Lightning 30B A3B	30B (3B active)	INT4 Q8	147 tok/s 114 tok/s	~1.6s ~1.6s
#153	Qwen3-32B	32B	INT4 Q8	33 tok/s 19 tok/s	~11.8s ~11.8s
#154	Qwen2.5-72B	72B	INT4	15 tok/s	~28.6s
#155	Mistral Large 3	41B (675B active)	INT4	2 tok/s	~237.9s
#156	GLM-4	32B	INT4 Q8	26 tok/s 16 tok/s	~11.8s ~12.8s
#157	DeepSeek-R1 70B	70B	INT4	17 tok/s	~25.6s
#158	GLM-4V	9B	INT4 Q8 FP16	69 tok/s 49 tok/s 30 tok/s	~3.4s ~3.4s ~3.4s
#159	Gemma 3 27B	27B	INT4 Q8	39 tok/s 23 tok/s	~10.0s ~10.0s
#160	Phi-4	14B	INT4 Q8 FP16	66 tok/s 41 tok/s 22 tok/s	~5.2s ~5.2s ~5.3s
#161	Llama 3.3 70B	70B	INT4	16 tok/s	~25.6s
#162	DeepSeek-R1 32B	32B	INT4 Q8	34 tok/s 20 tok/s	~11.8s ~11.8s
#163	Hunyuan Standard	52B (389B active)	INT4	3 tok/s	~137.4s
#164	Qwen2.5-32B	32B	INT4 Q8	33 tok/s 19 tok/s	~11.8s ~11.8s
#168	Magistral Small	24B	INT4 Q8	33 tok/s 22 tok/s	~8.9s ~8.9s
#169	OLMo 3 7B Think	7B	INT4 Q8 FP16	81 tok/s 60 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
#170	OLMo 3.1 32B Instruct	32B	INT4 Q8	26 tok/s 16 tok/s	~11.8s ~12.8s
#171	Llama 3.1 70B	70B	INT4	16 tok/s	~25.6s
#172	DeepSeek-R1 14B	14B	INT4 Q8 FP16	68 tok/s 41 tok/s 22 tok/s	~5.2s ~5.2s ~5.3s
#173	Ministral 3 14B	14B	INT4 Q8 FP16	50 tok/s 34 tok/s 20 tok/s	~5.2s ~5.3s ~5.3s
#177	Gemma 3 12B	12B	INT4 Q8 FP16	74 tok/s 46 tok/s 26 tok/s	~4.5s ~4.5s ~4.5s
#179	Qwen2-72B	72B	INT4	15 tok/s	~28.6s
#180	OLMo 3 32B Base	32B	INT4 Q8	26 tok/s 16 tok/s	~11.8s ~12.8s
#181	Mistral-Small-2501	24B	INT4 Q8 FP16	43 tok/s 25 tok/s 12 tok/s	~8.9s ~8.9s ~10.5s
#182	Ministral 3 8B	8B	INT4 Q8 FP16	75 tok/s 54 tok/s 33 tok/s	~3.0s ~3.0s ~3.0s
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	194 tok/s 154 tok/s 106 tok/s	~829 ms ~830 ms ~833 ms
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	194 tok/s 154 tok/s 106 tok/s	~829 ms ~830 ms ~833 ms
#185	Qwen2.5-14B	14B	INT4 Q8 FP16	66 tok/s 41 tok/s 22 tok/s	~5.2s ~5.2s ~5.3s
#186	Gemma 2 27B	27B	INT4 Q8	39 tok/s 23 tok/s	~10.0s ~10.0s
#187	OLMo 3.1 32B Think	32B	INT4 Q8	26 tok/s 16 tok/s	~11.8s ~12.8s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	129 tok/s 90 tok/s 54 tok/s	~2.0s ~2.0s ~2.0s
#189	Qwen3-4B	4B	INT4 Q8 FP16	147 tok/s 105 tok/s 65 tok/s	~1.6s ~1.6s ~1.6s
#190	DeepSeek-R1 8B	8B	INT4 Q8 FP16	101 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#191	Llama 3 70B	70B	INT4	16 tok/s	~25.6s
#193	Ministral 3 3B	3B	INT4 Q8 FP16	124 tok/s 101 tok/s 70 tok/s	~1.2s ~1.2s ~1.2s
#195	Qwen3-14B	14B	INT4 Q8 FP16	66 tok/s 41 tok/s 22 tok/s	~5.2s ~5.2s ~5.3s
#196	Kimi Linear 48B A3B Instruct	48B (3B active)	INT4	59 tok/s	~1.7s
#197	Gemma 2 9B	9B	INT4 Q8 FP16	91 tok/s 59 tok/s 33 tok/s	~3.4s ~3.4s ~3.4s
#198	Qwen2.5-7B	7B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
#199	Qwen3-8B	8B	INT4 Q8 FP16	98 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#200	Gemma 3 4B	4B	INT4 Q8 FP16	147 tok/s 105 tok/s 65 tok/s	~1.6s ~1.6s ~1.6s
#201	DeepSeek-R1 7B	7B	INT4 Q8 FP16	109 tok/s 72 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
#202	Llama 3.1 8B	8B	INT4 Q8 FP16	98 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	204 tok/s 166 tok/s 117 tok/s	~710 ms ~711 ms ~713 ms
#205	Command R	35B	INT4 Q8	24 tok/s 15 tok/s	~12.8s ~14.0s
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	268 tok/s 257 tok/s 233 tok/s	~115 ms ~116 ms ~116 ms
#208	Llama 3 8B	8B	INT4 Q8 FP16	98 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	259 tok/s 235 tok/s 195 tok/s	~252 ms ~253 ms ~253 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	150 tok/s 109 tok/s 68 tok/s	~1.5s ~1.5s ~1.5s
#211	Mixtral-8x7B-v0.1	46.7B (7B active)	INT4 Q8	97 tok/s 60 tok/s	~3.1s ~3.6s
#212	Ministral-8B-2410	8B	INT4 Q8 FP16	98 tok/s 65 tok/s 37 tok/s	~3.0s ~3.0s ~3.0s
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	212 tok/s 175 tok/s 126 tok/s	~628 ms ~629 ms ~631 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	167 tok/s 125 tok/s 81 tok/s	~1.2s ~1.2s ~1.2s
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	247 tok/s 218 tok/s 174 tok/s	~340 ms ~340 ms ~341 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	150 tok/s 109 tok/s 68 tok/s	~1.5s ~1.5s ~1.5s
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	124 tok/s 109 tok/s 83 tok/s	~888 ms ~889 ms ~892 ms
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	105 tok/s 69 tok/s 40 tok/s	~2.8s ~2.8s ~2.8s
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	235 tok/s 203 tok/s 157 tok/s	~431 ms ~431 ms ~433 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	194 tok/s 154 tok/s 106 tok/s	~829 ms ~830 ms ~833 ms
#221	Phi-3-medium	14B	INT4 Q8 FP16	66 tok/s 41 tok/s 22 tok/s	~5.2s ~5.2s ~5.3s
#222	OLMo 3 7B Instruct	7B	INT4 Q8 FP16	81 tok/s 60 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
#223	Yi-34B	34B	INT4 Q8	25 tok/s 15 tok/s	~12.5s ~13.6s
#224	Qwen2-7B	7B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
#225	Phi-3-small	7B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
#226	Gemma 3 1B	1B	INT4 Q8 FP16	235 tok/s 203 tok/s 157 tok/s	~431 ms ~431 ms ~433 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	105 tok/s 69 tok/s 40 tok/s	~2.8s ~2.8s ~2.8s
#229	Gemma 1 7B	7B	INT4 Q8 FP16	81 tok/s 60 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
#230	Gemma 1 2B	2B	INT4 Q8 FP16	201 tok/s 159 tok/s 108 tok/s	~829 ms ~830 ms ~833 ms
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	89 tok/s 67 tok/s 42 tok/s	~2.3s ~2.3s ~2.3s
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	89 tok/s 67 tok/s 42 tok/s	~2.3s ~2.3s ~2.3s
#233	ChatGLM-6B	6B	INT4 Q8 FP16	89 tok/s 67 tok/s 42 tok/s	~2.3s ~2.3s ~2.3s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	266 tok/s 245 tok/s 207 tok/s	~210 ms ~210 ms ~211 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	266 tok/s 245 tok/s 207 tok/s	~210 ms ~210 ms ~211 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	264 tok/s 252 tok/s 226 tok/s	~128 ms ~128 ms ~129 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	264 tok/s 252 tok/s 226 tok/s	~128 ms ~128 ms ~129 ms
-	Falcon-1B	1B	INT4 Q8 FP16	242 tok/s 208 tok/s 159 tok/s	~430 ms ~431 ms ~433 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	235 tok/s 203 tok/s 157 tok/s	~431 ms ~431 ms ~433 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	212 tok/s 175 tok/s 126 tok/s	~628 ms ~629 ms ~631 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	212 tok/s 175 tok/s 126 tok/s	~628 ms ~629 ms ~631 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	204 tok/s 166 tok/s 117 tok/s	~710 ms ~711 ms ~713 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	201 tok/s 162 tok/s 113 tok/s	~747 ms ~748 ms ~751 ms
-	Falcon-3B	3B	INT4 Q8 FP16	173 tok/s 129 tok/s 82 tok/s	~1.2s ~1.2s ~1.2s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	171 tok/s 152 tok/s 118 tok/s	~549 ms ~550 ms ~552 ms
-	Phi-1	1.3B	INT4 Q8 FP16	171 tok/s 152 tok/s 118 tok/s	~549 ms ~550 ms ~552 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	171 tok/s 152 tok/s 118 tok/s	~549 ms ~550 ms ~552 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	168 tok/s 126 tok/s 81 tok/s	~1.2s ~1.2s ~1.2s
-	Falcon3-3B	3B	INT4 Q8 FP16	167 tok/s 125 tok/s 81 tok/s	~1.2s ~1.2s ~1.2s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	167 tok/s 125 tok/s 81 tok/s	~1.2s ~1.2s ~1.2s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4 Q8 FP16	152 tok/s 117 tok/s 73 tok/s	~1.5s ~1.5s ~1.6s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4 Q8 FP16	152 tok/s 117 tok/s 73 tok/s	~1.5s ~1.5s ~1.6s
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4 Q8	148 tok/s 115 tok/s	~1.5s ~1.5s
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4 Q8	148 tok/s 115 tok/s	~1.5s ~1.5s
-	Hy-MT2 30B A3B	30B (3B active)	INT4 Q8	147 tok/s 114 tok/s	~1.6s ~1.6s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	147 tok/s 105 tok/s 65 tok/s	~1.6s ~1.6s ~1.6s
-	Phi-2	2.7B	INT4 Q8 FP16	129 tok/s 107 tok/s 75 tok/s	~1.1s ~1.1s ~1.1s
-	MaLLaM-3B	3B	INT4 Q8 FP16	124 tok/s 101 tok/s 70 tok/s	~1.2s ~1.2s ~1.2s
-	SmolLM3 3B	3B	INT4 Q8 FP16	124 tok/s 101 tok/s 70 tok/s	~1.2s ~1.2s ~1.2s
-	AliceAI Foundation 80B A3B Base	80B (3B active)	INT4	117 tok/s	~2.3s
-	Falcon-7B	7B	INT4 Q8 FP16	111 tok/s 73 tok/s 42 tok/s	~2.7s ~2.7s ~2.7s
-	Falcon3-7B	7B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
-	Hy-MT2 7B	7B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~2.7s ~2.7s ~2.7s
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	105 tok/s 69 tok/s 40 tok/s	~2.8s ~2.8s ~2.8s
-	LensVLM 9B	9B	INT4 Q8 FP16	91 tok/s 59 tok/s 33 tok/s	~3.4s ~3.4s ~3.4s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8 FP16	91 tok/s 59 tok/s 33 tok/s	~3.4s ~3.4s ~3.4s
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	89 tok/s 67 tok/s 42 tok/s	~2.3s ~2.3s ~2.3s
-	Yi-6B	6B	INT4 Q8 FP16	89 tok/s 67 tok/s 42 tok/s	~2.3s ~2.3s ~2.3s
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4 Q8 FP16	85 tok/s 75 tok/s 56 tok/s	~1.4s ~1.4s ~1.4s
-	Falcon3-10B	10B	INT4 Q8 FP16	84 tok/s 54 tok/s 30 tok/s	~3.8s ~3.8s ~3.8s
-	Falcon2-11B	11B	INT4 Q8 FP16	82 tok/s 51 tok/s 28 tok/s	~4.1s ~4.1s ~4.2s
-	Hunyuan Lite	7B	INT4 Q8 FP16	81 tok/s 60 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
-	OLMo 3 7B Base	7B	INT4 Q8 FP16	81 tok/s 60 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
-	SEA-LION-7B	7.1B	INT4 Q8 FP16	81 tok/s 59 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8 FP16	81 tok/s 59 tok/s 37 tok/s	~2.7s ~2.7s ~2.7s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8 FP16	75 tok/s 54 tok/s 33 tok/s	~3.0s ~3.0s ~3.0s
-	Typhoon-2-8B	8B	INT4 Q8 FP16	75 tok/s 54 tok/s 33 tok/s	~3.0s ~3.0s ~3.0s
-	GLM-4-9B	9B	INT4 Q8 FP16	69 tok/s 49 tok/s 30 tok/s	~3.4s ~3.4s ~3.4s
-	GLM-4-9B-Chat	9B	INT4 Q8 FP16	69 tok/s 49 tok/s 30 tok/s	~3.4s ~3.4s ~3.4s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8 FP16	69 tok/s 49 tok/s 30 tok/s	~3.4s ~3.4s ~3.4s
-	Yi-9B	9B	INT4 Q8 FP16	69 tok/s 49 tok/s 30 tok/s	~3.4s ~3.4s ~3.4s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8 FP16	68 tok/s 49 tok/s 29 tok/s	~3.5s ~3.5s ~3.5s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8 FP16	68 tok/s 49 tok/s 29 tok/s	~3.5s ~3.5s ~3.5s
-	GreenMind-14B-R1	14B	INT4 Q8 FP16	50 tok/s 34 tok/s 20 tok/s	~5.2s ~5.3s ~5.3s
-	Falcon-40B	40B	INT4 Q8	28 tok/s 15 tok/s	~14.6s ~15.9s
-	Typhoon-2-70B	70B	INT4	12 tok/s	~30.1s
-	Kimi-Dev-72B	72B	INT4	11 tok/s	~30.9s

Apple M2 Ultra (64GB)

VRAM Scaling by Context Length

Best Models to Run on Apple M2 Ultra (64GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions