What are the best models to run on Apple M3 Max (48GB)?

Based on model rankings and 36 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~21 tok/s. Top Coding: Qwen3.5-27B (INT4) ~21 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~27 tok/s. Maximum Model Size: DeepSeek-R1 70B (INT4) ~8 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 M3 Max (48GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4 Q8	21 tok/s 11 tok/s	~16.5s ~18.0s
#48	Agnes 3.0 Flash	33B	INT4 Q8	17 tok/s 9 tok/s	~20.1s ~23.7s
#72	K2 Horizon MoVA 36B A4B	36B (4B active)	INT4 Q8	82 tok/s 52 tok/s	~3.2s ~3.8s
#83	Qwen3.5-27B	27B	INT4 Q8	21 tok/s 11 tok/s	~16.5s ~18.0s
#92	Sarvam-30B	32B (2.4B active)	INT4 Q8	106 tok/s 76 tok/s	~2.3s ~2.5s
#94	Qwen3.6 35B A3B	35B (3B active)	INT4 Q8	95 tok/s 63 tok/s	~2.6s ~3.1s
#96	Muse Glimmer 30B	30B	INT4 Q8	19 tok/s 10 tok/s	~18.3s ~19.9s
#100	Phi-4 Reasoning Plus	14B	INT4 Q8 FP16	27 tok/s 18 tok/s 10 tok/s	~8.7s ~8.7s ~9.5s
#101	MiMo V2 Flash	15B (309B active)	INT4 Q8 FP16	2 tok/s 1 tok/s 1 tok/s	~180.7s ~181.1s ~214.0s
#108	Gemma 4 26B A4B	25.2B (3.8B active)	INT4 Q8	88 tok/s 61 tok/s	~3.0s ~3.0s
#109	Gemma 4 12B	11.95B	INT4 Q8 FP16	31 tok/s 21 tok/s 12 tok/s	~7.4s ~7.5s ~7.5s
#113	Qwen3.5-9B	9B	INT4 Q8 FP16	53 tok/s 33 tok/s 17 tok/s	~5.6s ~5.6s ~5.7s
#114	Qwen3.5-35B-A3B	35B (3B active)	INT4 Q8	95 tok/s 63 tok/s	~2.6s ~3.1s
#118	K2 Horizon 7B	7B	INT4 Q8 FP16	65 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	18 tok/s 10 tok/s 5 tok/s	~17.7s ~17.8s ~17.9s
#129	Gemma 4 31B	30.7B	INT4 Q8	18 tok/s 10 tok/s	~18.7s ~20.4s
#131	Nemotron 3.5 Lightning	30B (3B active)	INT4 Q8	38 tok/s 31 tok/s	~2.6s ~3.1s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	97 tok/s 64 tok/s 36 tok/s	~2.6s ~2.6s ~2.6s
#133	Qwen3-30B-A3B	30B (3B active)	INT4 Q8	97 tok/s 67 tok/s	~2.6s ~2.8s
#136	GPT-OSS 20B	21B (3.6B active)	INT4 Q8	43 tok/s 37 tok/s	~2.8s ~2.8s
#146	Granite 4.2 8B	8B	INT4 Q8 FP16	59 tok/s 36 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#150	Gemma 4 E4B	8B	INT4 Q8 FP16	59 tok/s 36 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#151	Nemotron 3.5 Lightning 30B A3B	30B (3B active)	INT4 Q8	97 tok/s 67 tok/s	~2.6s ~2.8s
#153	Qwen3-32B	32B	INT4 Q8	18 tok/s 10 tok/s	~19.5s ~21.2s
#155	Mistral Large 3	41B (675B active)	INT4	1 tok/s	~428.7s
#156	GLM-4	32B	INT4	14 tok/s	~19.5s
#157	DeepSeek-R1 70B	70B	INT4	8 tok/s	~49.8s
#158	GLM-4V	9B	INT4 Q8 FP16	39 tok/s 27 tok/s 16 tok/s	~5.6s ~5.6s ~5.7s
#159	Gemma 3 27B	27B	INT4 Q8	21 tok/s 11 tok/s	~16.5s ~18.0s
#160	Phi-4	14B	INT4 Q8 FP16	37 tok/s 22 tok/s 11 tok/s	~8.7s ~8.7s ~9.5s
#161	Llama 3.3 70B	70B	INT4	8 tok/s	~49.8s
#162	DeepSeek-R1 32B	32B	INT4 Q8	18 tok/s 10 tok/s	~19.5s ~21.2s
#163	Hunyuan Standard	52B (389B active)	INT4	1 tok/s	~268.0s
#164	Qwen2.5-32B	32B	INT4 Q8	18 tok/s 10 tok/s	~19.5s ~21.2s
#168	Magistral Small	24B	INT4 Q8	17 tok/s 11 tok/s	~14.7s ~16.0s
#169	OLMo 3 7B Think	7B	INT4 Q8 FP16	47 tok/s 33 tok/s 20 tok/s	~4.4s ~4.4s ~4.4s
#170	OLMo 3.1 32B Instruct	32B	INT4	14 tok/s	~19.5s
#171	Llama 3.1 70B	70B	INT4	8 tok/s	~49.8s
#172	DeepSeek-R1 14B	14B	INT4 Q8 FP16	38 tok/s 22 tok/s 11 tok/s	~8.7s ~8.7s ~9.5s
#173	Ministral 3 14B	14B	INT4 Q8 FP16	27 tok/s 18 tok/s 10 tok/s	~8.7s ~8.7s ~9.5s
#177	Gemma 3 12B	12B	INT4 Q8 FP16	42 tok/s 25 tok/s 13 tok/s	~7.5s ~7.5s ~7.5s
#180	OLMo 3 32B Base	32B	INT4	14 tok/s	~19.5s
#181	Mistral-Small-2501	24B	INT4 Q8	23 tok/s 13 tok/s	~14.7s ~14.8s
#182	Ministral 3 8B	8B	INT4 Q8 FP16	43 tok/s 30 tok/s 17 tok/s	~5.0s ~5.0s ~5.0s
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	142 tok/s 103 tok/s 64 tok/s	~1.4s ~1.4s ~1.4s
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	142 tok/s 103 tok/s 64 tok/s	~1.4s ~1.4s ~1.4s
#185	Qwen2.5-14B	14B	INT4 Q8 FP16	37 tok/s 22 tok/s 11 tok/s	~8.7s ~8.7s ~9.5s
#186	Gemma 2 27B	27B	INT4 Q8	21 tok/s 11 tok/s	~16.5s ~18.0s
#187	OLMo 3.1 32B Think	32B	INT4	14 tok/s	~19.5s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	82 tok/s 53 tok/s 29 tok/s	~3.2s ~3.3s ~3.3s
#189	Qwen3-4B	4B	INT4 Q8 FP16	97 tok/s 64 tok/s 36 tok/s	~2.6s ~2.6s ~2.6s
#190	DeepSeek-R1 8B	8B	INT4 Q8 FP16	60 tok/s 37 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#191	Llama 3 70B	70B	INT4	8 tok/s	~49.8s
#193	Ministral 3 3B	3B	INT4 Q8 FP16	78 tok/s 61 tok/s 40 tok/s	~2.0s ~2.0s ~2.0s
#195	Qwen3-14B	14B	INT4 Q8 FP16	37 tok/s 22 tok/s 11 tok/s	~8.7s ~8.7s ~9.5s
#196	Kimi Linear 48B A3B Instruct	48B (3B active)	INT4	31 tok/s	~3.0s
#197	Gemma 2 9B	9B	INT4 Q8 FP16	53 tok/s 33 tok/s 17 tok/s	~5.6s ~5.6s ~5.7s
#198	Qwen2.5-7B	7B	INT4 Q8 FP16	65 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
#199	Qwen3-8B	8B	INT4 Q8 FP16	59 tok/s 36 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#200	Gemma 3 4B	4B	INT4 Q8 FP16	97 tok/s 64 tok/s 36 tok/s	~2.6s ~2.6s ~2.6s
#201	DeepSeek-R1 7B	7B	INT4 Q8 FP16	67 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
#202	Llama 3.1 8B	8B	INT4 Q8 FP16	59 tok/s 36 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	154 tok/s 114 tok/s 73 tok/s	~1.2s ~1.2s ~1.2s
#205	Command R	35B	INT4	13 tok/s	~21.3s
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	240 tok/s 222 tok/s 189 tok/s	~186 ms ~186 ms ~187 ms
#208	Llama 3 8B	8B	INT4 Q8 FP16	59 tok/s 36 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	227 tok/s 192 tok/s 144 tok/s	~413 ms ~414 ms ~415 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	100 tok/s 66 tok/s 38 tok/s	~2.5s ~2.5s ~2.5s
#211	Mixtral-8x7B-v0.1	46.7B (7B active)	INT4	58 tok/s	~5.1s
#212	Ministral-8B-2410	8B	INT4 Q8 FP16	59 tok/s 36 tok/s 20 tok/s	~5.0s ~5.0s ~5.0s
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	163 tok/s 123 tok/s 80 tok/s	~1.0s ~1.0s ~1.0s
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	115 tok/s 79 tok/s 47 tok/s	~2.0s ~2.0s ~2.0s
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	208 tok/s 171 tok/s 122 tok/s	~558 ms ~559 ms ~561 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	100 tok/s 66 tok/s 38 tok/s	~2.5s ~2.5s ~2.5s
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	78 tok/s 67 tok/s 48 tok/s	~1.5s ~1.5s ~1.5s
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	63 tok/s 39 tok/s 21 tok/s	~4.6s ~4.6s ~4.6s
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	192 tok/s 153 tok/s 106 tok/s	~708 ms ~710 ms ~713 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	142 tok/s 103 tok/s 64 tok/s	~1.4s ~1.4s ~1.4s
#221	Phi-3-medium	14B	INT4 Q8 FP16	37 tok/s 22 tok/s 11 tok/s	~8.7s ~8.7s ~9.5s
#222	OLMo 3 7B Instruct	7B	INT4 Q8 FP16	47 tok/s 33 tok/s 20 tok/s	~4.4s ~4.4s ~4.4s
#223	Yi-34B	34B	INT4	13 tok/s	~20.7s
#224	Qwen2-7B	7B	INT4 Q8 FP16	65 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
#225	Phi-3-small	7B	INT4 Q8 FP16	65 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
#226	Gemma 3 1B	1B	INT4 Q8 FP16	192 tok/s 153 tok/s 106 tok/s	~708 ms ~710 ms ~713 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	63 tok/s 39 tok/s 21 tok/s	~4.6s ~4.6s ~4.6s
#229	Gemma 1 7B	7B	INT4 Q8 FP16	47 tok/s 33 tok/s 20 tok/s	~4.4s ~4.4s ~4.4s
#230	Gemma 1 2B	2B	INT4 Q8 FP16	151 tok/s 108 tok/s 66 tok/s	~1.4s ~1.4s ~1.4s
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	52 tok/s 37 tok/s 23 tok/s	~3.8s ~3.8s ~3.8s
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	52 tok/s 37 tok/s 23 tok/s	~3.8s ~3.8s ~3.8s
#233	ChatGLM-6B	6B	INT4 Q8 FP16	52 tok/s 37 tok/s 23 tok/s	~3.8s ~3.8s ~3.8s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	237 tok/s 205 tok/s 158 tok/s	~342 ms ~343 ms ~345 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	237 tok/s 205 tok/s 158 tok/s	~342 ms ~343 ms ~345 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	233 tok/s 215 tok/s 181 tok/s	~207 ms ~207 ms ~208 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	233 tok/s 215 tok/s 181 tok/s	~207 ms ~207 ms ~208 ms
-	Falcon-1B	1B	INT4 Q8 FP16	201 tok/s 158 tok/s 108 tok/s	~708 ms ~710 ms ~713 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	192 tok/s 153 tok/s 106 tok/s	~708 ms ~710 ms ~713 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	163 tok/s 123 tok/s 80 tok/s	~1.0s ~1.0s ~1.0s
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	163 tok/s 123 tok/s 80 tok/s	~1.0s ~1.0s ~1.0s
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	154 tok/s 114 tok/s 73 tok/s	~1.2s ~1.2s ~1.2s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	150 tok/s 111 tok/s 70 tok/s	~1.2s ~1.2s ~1.2s
-	Falcon-3B	3B	INT4 Q8 FP16	121 tok/s 82 tok/s 47 tok/s	~2.0s ~2.0s ~2.0s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	119 tok/s 101 tok/s 73 tok/s	~905 ms ~907 ms ~911 ms
-	Phi-1	1.3B	INT4 Q8 FP16	119 tok/s 101 tok/s 73 tok/s	~905 ms ~907 ms ~911 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	119 tok/s 101 tok/s 73 tok/s	~905 ms ~907 ms ~911 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	117 tok/s 80 tok/s 47 tok/s	~2.0s ~2.0s ~2.0s
-	Falcon3-3B	3B	INT4 Q8 FP16	115 tok/s 79 tok/s 47 tok/s	~2.0s ~2.0s ~2.0s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	115 tok/s 79 tok/s 47 tok/s	~2.0s ~2.0s ~2.0s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4 Q8	101 tok/s 73 tok/s	~2.4s ~2.4s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4 Q8	101 tok/s 73 tok/s	~2.4s ~2.4s
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4 Q8	98 tok/s 67 tok/s	~2.5s ~2.8s
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4 Q8	98 tok/s 67 tok/s	~2.5s ~2.8s
-	Hy-MT2 30B A3B	30B (3B active)	INT4 Q8	97 tok/s 67 tok/s	~2.6s ~2.8s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	97 tok/s 64 tok/s 36 tok/s	~2.6s ~2.6s ~2.6s
-	Phi-2	2.7B	INT4 Q8 FP16	82 tok/s 65 tok/s 43 tok/s	~1.8s ~1.8s ~1.8s
-	MaLLaM-3B	3B	INT4 Q8 FP16	78 tok/s 61 tok/s 40 tok/s	~2.0s ~2.0s ~2.0s
-	SmolLM3 3B	3B	INT4 Q8 FP16	78 tok/s 61 tok/s 40 tok/s	~2.0s ~2.0s ~2.0s
-	Falcon-7B	7B	INT4 Q8 FP16	68 tok/s 42 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
-	Falcon3-7B	7B	INT4 Q8 FP16	65 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
-	Hy-MT2 7B	7B	INT4 Q8 FP16	65 tok/s 41 tok/s 22 tok/s	~4.4s ~4.4s ~4.4s
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	63 tok/s 39 tok/s 21 tok/s	~4.6s ~4.6s ~4.6s
-	LensVLM 9B	9B	INT4 Q8 FP16	53 tok/s 33 tok/s 17 tok/s	~5.6s ~5.6s ~5.7s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8 FP16	53 tok/s 33 tok/s 17 tok/s	~5.6s ~5.6s ~5.7s
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	52 tok/s 37 tok/s 23 tok/s	~3.8s ~3.8s ~3.8s
-	Yi-6B	6B	INT4 Q8 FP16	52 tok/s 37 tok/s 23 tok/s	~3.8s ~3.8s ~3.8s
-	Falcon3-10B	10B	INT4 Q8 FP16	49 tok/s 30 tok/s 16 tok/s	~6.2s ~6.2s ~6.3s
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4 Q8 FP16	49 tok/s 43 tok/s 28 tok/s	~2.3s ~2.3s ~2.7s
-	Falcon2-11B	11B	INT4 Q8 FP16	47 tok/s 28 tok/s 15 tok/s	~6.8s ~6.9s ~6.9s
-	Hunyuan Lite	7B	INT4 Q8 FP16	47 tok/s 33 tok/s 20 tok/s	~4.4s ~4.4s ~4.4s
-	OLMo 3 7B Base	7B	INT4 Q8 FP16	47 tok/s 33 tok/s 20 tok/s	~4.4s ~4.4s ~4.4s
-	SEA-LION-7B	7.1B	INT4 Q8 FP16	46 tok/s 33 tok/s 19 tok/s	~4.5s ~4.5s ~4.5s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8 FP16	46 tok/s 33 tok/s 19 tok/s	~4.5s ~4.5s ~4.5s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8 FP16	43 tok/s 30 tok/s 17 tok/s	~5.0s ~5.0s ~5.0s
-	Typhoon-2-8B	8B	INT4 Q8 FP16	43 tok/s 30 tok/s 17 tok/s	~5.0s ~5.0s ~5.0s
-	GLM-4-9B	9B	INT4 Q8 FP16	39 tok/s 27 tok/s 16 tok/s	~5.6s ~5.6s ~5.7s
-	GLM-4-9B-Chat	9B	INT4 Q8 FP16	39 tok/s 27 tok/s 16 tok/s	~5.6s ~5.6s ~5.7s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8 FP16	39 tok/s 27 tok/s 16 tok/s	~5.6s ~5.6s ~5.7s
-	Yi-9B	9B	INT4 Q8 FP16	39 tok/s 27 tok/s 16 tok/s	~5.6s ~5.6s ~5.7s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8 FP16	38 tok/s 26 tok/s 15 tok/s	~5.8s ~5.8s ~5.8s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8 FP16	38 tok/s 26 tok/s 15 tok/s	~5.8s ~5.8s ~5.8s
-	GreenMind-14B-R1	14B	INT4 Q8 FP16	27 tok/s 18 tok/s 10 tok/s	~8.7s ~8.7s ~9.5s
-	Falcon-40B	40B	INT4	15 tok/s	~24.3s

Apple M3 Max (48GB)

VRAM Scaling by Context Length

Best Models to Run on Apple M3 Max (48GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions