What are the best models to run on NVIDIA RTX 4000 Ada Generation (20GB)?

Based on model rankings and 20 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~19 tok/s. Top Coding: Qwen3.5-27B (INT4) ~19 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~25 tok/s. Maximum Model Size: K2 Horizon MoVA 36B A4B (INT4) ~72 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 NVIDIA RTX 4000 Ada Generation (20GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4	19 tok/s	~2.8s
#48	Agnes 3.0 Flash	33B	INT4	15 tok/s	~3.7s
#72	K2 Horizon MoVA 36B A4B	36B (4B active)	INT4	72 tok/s	~645 ms
#83	Qwen3.5-27B	27B	INT4	19 tok/s	~2.8s
#92	Sarvam-30B	32B (2.4B active)	INT4	100 tok/s	~417 ms
#94	Qwen3.6 35B A3B	35B (3B active)	INT4	84 tok/s	~525 ms
#96	Muse Glimmer 30B	30B	INT4	16 tok/s	~3.4s
#100	Phi-4 Reasoning Plus	14B	INT4 Q8	25 tok/s 16 tok/s	~1.5s ~1.6s
#101	MiMo V2 Flash	15B (309B active)	INT4	2 tok/s	~30.4s
#108	Gemma 4 26B A4B	25.2B (3.8B active)	INT4	86 tok/s	~503 ms
#109	Gemma 4 12B	11.95B	INT4 Q8	28 tok/s 18 tok/s	~1.3s ~1.4s
#113	Qwen3.5-9B	9B	INT4 Q8 FP16	50 tok/s 30 tok/s 14 tok/s	~952 ms ~966 ms ~1.2s
#114	Qwen3.5-35B-A3B	35B (3B active)	INT4	84 tok/s	~525 ms
#118	K2 Horizon 7B	7B	INT4 Q8 FP16	62 tok/s 38 tok/s 20 tok/s	~747 ms ~757 ms ~780 ms
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	16 tok/s 9 tok/s 5 tok/s	~3.0s ~3.0s ~3.1s
#129	Gemma 4 31B	30.7B	INT4	16 tok/s	~3.4s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	95 tok/s 61 tok/s 34 tok/s	~440 ms ~446 ms ~459 ms
#133	Qwen3-30B-A3B	30B (3B active)	INT4	90 tok/s	~477 ms
#136	GPT-OSS 20B	21B (3.6B active)	INT4	38 tok/s	~525 ms
#146	Granite 4.2 8B	8B	INT4 Q8 FP16	56 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~961 ms
#150	Gemma 4 E4B	8B	INT4 Q8 FP16	56 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~961 ms
#151	Nemotron 3.5 Lightning 30B A3B	30B (3B active)	INT4	90 tok/s	~477 ms
#153	Qwen3-32B	32B	INT4	15 tok/s	~3.6s
#158	GLM-4V	9B	INT4 Q8	36 tok/s 25 tok/s	~960 ms ~973 ms
#159	Gemma 3 27B	27B	INT4	19 tok/s	~2.8s
#160	Phi-4	14B	INT4 Q8	34 tok/s 20 tok/s	~1.5s ~1.5s
#162	DeepSeek-R1 32B	32B	INT4	16 tok/s	~3.6s
#164	Qwen2.5-32B	32B	INT4	15 tok/s	~3.6s
#168	Magistral Small	24B	INT4	15 tok/s	~2.7s
#169	OLMo 3 7B Think	7B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~753 ms ~763 ms ~852 ms
#172	DeepSeek-R1 14B	14B	INT4 Q8	35 tok/s 20 tok/s	~1.5s ~1.5s
#173	Ministral 3 14B	14B	INT4 Q8	25 tok/s 16 tok/s	~1.5s ~1.6s
#177	Gemma 3 12B	12B	INT4 Q8	39 tok/s 23 tok/s	~1.3s ~1.3s
#181	Mistral-Small-2501	24B	INT4	21 tok/s	~2.5s
#182	Ministral 3 8B	8B	INT4 Q8 FP16	40 tok/s 27 tok/s 14 tok/s	~856 ms ~867 ms ~1.0s
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	146 tok/s 102 tok/s 61 tok/s	~237 ms ~240 ms ~246 ms
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	146 tok/s 102 tok/s 61 tok/s	~237 ms ~240 ms ~246 ms
#185	Qwen2.5-14B	14B	INT4 Q8	34 tok/s 20 tok/s	~1.5s ~1.5s
#186	Gemma 2 27B	27B	INT4	19 tok/s	~2.8s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	80 tok/s 50 tok/s 27 tok/s	~552 ms ~560 ms ~577 ms
#189	Qwen3-4B	4B	INT4 Q8 FP16	95 tok/s 61 tok/s 34 tok/s	~440 ms ~446 ms ~459 ms
#190	DeepSeek-R1 8B	8B	INT4 Q8 FP16	57 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~960 ms
#193	Ministral 3 3B	3B	INT4 Q8 FP16	75 tok/s 58 tok/s 37 tok/s	~343 ms ~347 ms ~356 ms
#195	Qwen3-14B	14B	INT4 Q8	34 tok/s 20 tok/s	~1.5s ~1.5s
#197	Gemma 2 9B	9B	INT4 Q8 FP16	50 tok/s 30 tok/s 14 tok/s	~952 ms ~966 ms ~1.2s
#198	Qwen2.5-7B	7B	INT4 Q8 FP16	62 tok/s 38 tok/s 20 tok/s	~747 ms ~757 ms ~780 ms
#199	Qwen3-8B	8B	INT4 Q8 FP16	56 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~961 ms
#200	Gemma 3 4B	4B	INT4 Q8 FP16	95 tok/s 61 tok/s 34 tok/s	~440 ms ~446 ms ~459 ms
#201	DeepSeek-R1 7B	7B	INT4 Q8 FP16	64 tok/s 38 tok/s 20 tok/s	~746 ms ~757 ms ~779 ms
#202	Llama 3.1 8B	8B	INT4 Q8 FP16	56 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~961 ms
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	160 tok/s 114 tok/s 70 tok/s	~203 ms ~206 ms ~211 ms
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	273 tok/s 248 tok/s 204 tok/s	~38 ms ~38 ms ~39 ms
#208	Llama 3 8B	8B	INT4 Q8 FP16	56 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~961 ms
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	254 tok/s 208 tok/s 148 tok/s	~76 ms ~77 ms ~79 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	98 tok/s 63 tok/s 35 tok/s	~420 ms ~425 ms ~438 ms
#212	Ministral-8B-2410	8B	INT4 Q8 FP16	56 tok/s 34 tok/s 17 tok/s	~848 ms ~860 ms ~961 ms
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	171 tok/s 124 tok/s 77 tok/s	~181 ms ~183 ms ~188 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	115 tok/s 76 tok/s 44 tok/s	~338 ms ~342 ms ~352 ms
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	229 tok/s 180 tok/s 123 tok/s	~100 ms ~101 ms ~104 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	98 tok/s 63 tok/s 35 tok/s	~420 ms ~425 ms ~438 ms
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	75 tok/s 64 tok/s 45 tok/s	~258 ms ~261 ms ~267 ms
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	60 tok/s 36 tok/s 19 tok/s	~777 ms ~787 ms ~811 ms
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	207 tok/s 159 tok/s 105 tok/s	~125 ms ~127 ms ~130 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	146 tok/s 102 tok/s 61 tok/s	~237 ms ~240 ms ~246 ms
#221	Phi-3-medium	14B	INT4 Q8	34 tok/s 20 tok/s	~1.5s ~1.5s
#222	OLMo 3 7B Instruct	7B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~753 ms ~763 ms ~852 ms
#224	Qwen2-7B	7B	INT4 Q8 FP16	62 tok/s 38 tok/s 20 tok/s	~747 ms ~757 ms ~780 ms
#225	Phi-3-small	7B	INT4 Q8 FP16	62 tok/s 38 tok/s 20 tok/s	~747 ms ~757 ms ~780 ms
#226	Gemma 3 1B	1B	INT4 Q8 FP16	207 tok/s 159 tok/s 105 tok/s	~125 ms ~127 ms ~130 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	60 tok/s 36 tok/s 19 tok/s	~777 ms ~787 ms ~811 ms
#229	Gemma 1 7B	7B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~753 ms ~763 ms ~852 ms
#230	Gemma 1 2B	2B	INT4 Q8 FP16	156 tok/s 107 tok/s 63 tok/s	~236 ms ~239 ms ~246 ms
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	49 tok/s 35 tok/s 21 tok/s	~651 ms ~659 ms ~679 ms
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	49 tok/s 35 tok/s 21 tok/s	~651 ms ~659 ms ~679 ms
#233	ChatGLM-6B	6B	INT4 Q8 FP16	49 tok/s 35 tok/s 21 tok/s	~651 ms ~659 ms ~679 ms
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	269 tok/s 225 tok/s 164 tok/s	~64 ms ~65 ms ~66 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	269 tok/s 225 tok/s 164 tok/s	~64 ms ~65 ms ~66 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	263 tok/s 238 tok/s 193 tok/s	~42 ms ~42 ms ~43 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	263 tok/s 238 tok/s 193 tok/s	~42 ms ~42 ms ~43 ms
-	Falcon-1B	1B	INT4 Q8 FP16	219 tok/s 165 tok/s 107 tok/s	~125 ms ~127 ms ~130 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	207 tok/s 159 tok/s 105 tok/s	~125 ms ~127 ms ~130 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	171 tok/s 124 tok/s 77 tok/s	~181 ms ~183 ms ~188 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	171 tok/s 124 tok/s 77 tok/s	~181 ms ~183 ms ~188 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	160 tok/s 114 tok/s 70 tok/s	~203 ms ~206 ms ~211 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	155 tok/s 110 tok/s 67 tok/s	~214 ms ~216 ms ~222 ms
-	Falcon-3B	3B	INT4 Q8 FP16	122 tok/s 79 tok/s 45 tok/s	~337 ms ~342 ms ~352 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	120 tok/s 100 tok/s 70 tok/s	~161 ms ~163 ms ~167 ms
-	Phi-1	1.3B	INT4 Q8 FP16	120 tok/s 100 tok/s 70 tok/s	~161 ms ~163 ms ~167 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	120 tok/s 100 tok/s 70 tok/s	~161 ms ~163 ms ~167 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	117 tok/s 77 tok/s 44 tok/s	~338 ms ~342 ms ~352 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	115 tok/s 76 tok/s 44 tok/s	~338 ms ~342 ms ~352 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	115 tok/s 76 tok/s 44 tok/s	~338 ms ~342 ms ~352 ms
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	100 tok/s	~412 ms
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	100 tok/s	~412 ms
-	Spark X2.5 4B	4B	INT4 Q8 FP16	95 tok/s 61 tok/s 34 tok/s	~440 ms ~446 ms ~459 ms
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4	91 tok/s	~469 ms
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4	91 tok/s	~469 ms
-	Hy-MT2 30B A3B	30B (3B active)	INT4	90 tok/s	~477 ms
-	Phi-2	2.7B	INT4 Q8 FP16	79 tok/s 62 tok/s 40 tok/s	~312 ms ~316 ms ~324 ms
-	MaLLaM-3B	3B	INT4 Q8 FP16	75 tok/s 58 tok/s 37 tok/s	~343 ms ~347 ms ~356 ms
-	SmolLM3 3B	3B	INT4 Q8 FP16	75 tok/s 58 tok/s 37 tok/s	~343 ms ~347 ms ~356 ms
-	Falcon-7B	7B	INT4 Q8 FP16	65 tok/s 39 tok/s 21 tok/s	~746 ms ~756 ms ~779 ms
-	Falcon3-7B	7B	INT4 Q8 FP16	62 tok/s 38 tok/s 20 tok/s	~747 ms ~757 ms ~780 ms
-	Hy-MT2 7B	7B	INT4 Q8 FP16	62 tok/s 38 tok/s 20 tok/s	~747 ms ~757 ms ~780 ms
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	60 tok/s 36 tok/s 19 tok/s	~777 ms ~787 ms ~811 ms
-	LensVLM 9B	9B	INT4 Q8 FP16	50 tok/s 30 tok/s 14 tok/s	~952 ms ~966 ms ~1.2s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8 FP16	50 tok/s 30 tok/s 14 tok/s	~952 ms ~966 ms ~1.2s
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	49 tok/s 35 tok/s 21 tok/s	~651 ms ~659 ms ~679 ms
-	Yi-6B	6B	INT4 Q8 FP16	49 tok/s 35 tok/s 21 tok/s	~651 ms ~659 ms ~679 ms
-	Falcon3-10B	10B	INT4 Q8	46 tok/s 27 tok/s	~1.1s ~1.1s
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4 Q8	46 tok/s 36 tok/s	~407 ms ~480 ms
-	Falcon2-11B	11B	INT4 Q8	44 tok/s 26 tok/s	~1.2s ~1.2s
-	Hunyuan Lite	7B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~753 ms ~763 ms ~852 ms
-	OLMo 3 7B Base	7B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~753 ms ~763 ms ~852 ms
-	SEA-LION-7B	7.1B	INT4 Q8 FP16	44 tok/s 30 tok/s 17 tok/s	~763 ms ~773 ms ~863 ms
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8 FP16	44 tok/s 30 tok/s 17 tok/s	~763 ms ~773 ms ~863 ms
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8 FP16	40 tok/s 27 tok/s 14 tok/s	~856 ms ~867 ms ~1.0s
-	Typhoon-2-8B	8B	INT4 Q8 FP16	40 tok/s 27 tok/s 14 tok/s	~856 ms ~867 ms ~1.0s
-	GLM-4-9B	9B	INT4 Q8	36 tok/s 25 tok/s	~960 ms ~973 ms
-	GLM-4-9B-Chat	9B	INT4 Q8	36 tok/s 25 tok/s	~960 ms ~973 ms
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	36 tok/s 25 tok/s	~960 ms ~973 ms
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8	36 tok/s 24 tok/s	~980 ms ~993 ms
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8	36 tok/s 24 tok/s	~980 ms ~993 ms
-	Yi-9B	9B	INT4 Q8	36 tok/s 25 tok/s	~960 ms ~973 ms
-	GreenMind-14B-R1	14B	INT4 Q8	25 tok/s 16 tok/s	~1.5s ~1.6s

NVIDIA RTX 4000 Ada Generation (20GB)

VRAM Scaling by Context Length

Best Models to Run on NVIDIA RTX 4000 Ada Generation (20GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions