What are the best models to run on NVIDIA RTX 3090 (24GB)?

Based on model rankings and 24 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~45 tok/s. Top Coding: Qwen3.5-27B (INT4) ~45 tok/s. Optimal Daily Driver: MiMo V2 Flash (INT4) ~5 tok/s. Maximum Model Size: Falcon-40B (INT4) ~31 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 3090 (24GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4	45 tok/s	~2.1s
#49	Agnes 3.0 Flash	33B	INT4	38 tok/s	~2.5s
#72	K2 Horizon MoVA 36B A4B	36B (4B active)	INT4	147 tok/s	~445 ms
#80	Qwen3.5-27B	27B	INT4	45 tok/s	~2.1s
#84	MiMo V2 Flash	15B (309B active)	INT4 Q8	5 tok/s 2 tok/s	~22.6s ~24.7s
#87	Sarvam-30B	32B (2.4B active)	INT4	188 tok/s	~288 ms
#90	Muse Glimmer 30B	30B	INT4	41 tok/s	~2.3s
#95	Qwen3.6 35B A3B	35B (3B active)	INT4	165 tok/s	~363 ms
#96	Phi-4 Reasoning Plus	14B	INT4 Q8	59 tok/s 40 tok/s	~1.1s ~1.1s
#102	Gemma 4 26B A4B	25.2B (3.8B active)	INT4	162 tok/s	~376 ms
#110	Qwen3.5-35B-A3B	35B (3B active)	INT4	165 tok/s	~363 ms
#117	Qwen3.5-9B	9B	INT4 Q8 FP16	108 tok/s 69 tok/s 37 tok/s	~710 ms ~715 ms ~788 ms
#119	K2 Horizon 7B	7B	INT4 Q8 FP16	127 tok/s 85 tok/s 48 tok/s	~557 ms ~561 ms ~570 ms
#124	Gemma 4 12B	11.95B	INT4 Q8	66 tok/s 46 tok/s	~940 ms ~946 ms
#127	Gemma 4 31B	30.7B	INT4	41 tok/s	~2.3s
#130	Nemotron 3.5 Lightning	30B (3B active)	INT4	76 tok/s	~363 ms
#131	Qwen3.5-4B	4B	INT4 Q8 FP16	175 tok/s 125 tok/s 77 tok/s	~329 ms ~331 ms ~336 ms
#133	GPT-OSS 20B	21B (3.6B active)	INT4	88 tok/s	~357 ms
#139	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	39 tok/s 24 tok/s 12 tok/s	~2.2s ~2.2s ~2.3s
#143	Qwen3-30B-A3B	30B (3B active)	INT4	175 tok/s	~328 ms
#145	Granite 4.2 8B	8B	INT4 Q8 FP16	117 tok/s 76 tok/s 43 tok/s	~633 ms ~637 ms ~647 ms
#150	Gemma 4 E4B	8B	INT4 Q8 FP16	117 tok/s 76 tok/s 43 tok/s	~633 ms ~637 ms ~647 ms
#152	Nemotron 3.5 Lightning 30B A3B	30B (3B active)	INT4	175 tok/s	~328 ms
#155	GLM-4	32B	INT4	28 tok/s	~2.9s
#157	GLM-4V	9B	INT4 Q8 FP16	81 tok/s 58 tok/s 33 tok/s	~713 ms ~718 ms ~791 ms
#157	Phi-4	14B	INT4 Q8	78 tok/s 48 tok/s	~1.1s ~1.1s
#157	Gemma 3 27B	27B	INT4	45 tok/s	~2.1s
#162	Qwen3-32B	32B	INT4	39 tok/s	~2.4s
#163	Qwen2.5-32B	32B	INT4	39 tok/s	~2.4s
#164	DeepSeek-R1 32B	32B	INT4	40 tok/s	~2.4s
#166	Magistral Small	24B	INT4	39 tok/s	~1.9s
#167	OLMo 3.1 32B Instruct	32B	INT4	28 tok/s	~2.9s
#171	Ministral 3 14B	14B	INT4 Q8	59 tok/s 40 tok/s	~1.1s ~1.1s
#172	DeepSeek-R1 14B	14B	INT4 Q8	80 tok/s 49 tok/s	~1.1s ~1.1s
#175	Gemma 3 12B	12B	INT4 Q8 FP16	87 tok/s 55 tok/s 27 tok/s	~940 ms ~946 ms ~1.1s
#179	Mistral-Small-2501	24B	INT4	50 tok/s	~1.8s
#180	OLMo 3 32B Base	32B	INT4	28 tok/s	~2.9s
#181	Ministral 3 8B	8B	INT4 Q8 FP16	88 tok/s 64 tok/s 37 tok/s	~635 ms ~640 ms ~705 ms
#182	OLMo 3 7B Think	7B	INT4 Q8 FP16	96 tok/s 70 tok/s 43 tok/s	~559 ms ~563 ms ~572 ms
#183	Gemma 2 27B	27B	INT4	45 tok/s	~2.1s
#184	Qwen3.5-2B	2B	INT4 Q8 FP16	233 tok/s 184 tok/s 126 tok/s	~178 ms ~179 ms ~181 ms
#185	Qwen2.5-14B	14B	INT4 Q8	78 tok/s 48 tok/s	~1.1s ~1.1s
#186	MiniCPM5-2B	2B	INT4 Q8 FP16	233 tok/s 184 tok/s 126 tok/s	~178 ms ~179 ms ~181 ms
#187	OLMo 3.1 32B Think	32B	INT4	28 tok/s	~2.9s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	154 tok/s 106 tok/s 63 tok/s	~413 ms ~415 ms ~422 ms
#191	Ministral 3 3B	3B	INT4 Q8 FP16	148 tok/s 120 tok/s 83 tok/s	~255 ms ~256 ms ~260 ms
#192	DeepSeek-R1 8B	8B	INT4 Q8 FP16	119 tok/s 77 tok/s 43 tok/s	~632 ms ~637 ms ~647 ms
#194	Gemma 2 9B	9B	INT4 Q8 FP16	108 tok/s 69 tok/s 37 tok/s	~710 ms ~715 ms ~788 ms
#196	Qwen2.5-7B	7B	INT4 Q8 FP16	127 tok/s 85 tok/s 48 tok/s	~557 ms ~561 ms ~570 ms
#197	Gemma 3 4B	4B	INT4 Q8 FP16	175 tok/s 125 tok/s 77 tok/s	~329 ms ~331 ms ~336 ms
#198	Qwen3-4B	4B	INT4 Q8 FP16	175 tok/s 125 tok/s 77 tok/s	~329 ms ~331 ms ~336 ms
#199	DeepSeek-R1 7B	7B	INT4 Q8 FP16	130 tok/s 86 tok/s 49 tok/s	~557 ms ~561 ms ~569 ms
#200	Llama 3.1 8B	8B	INT4 Q8 FP16	117 tok/s 76 tok/s 43 tok/s	~633 ms ~637 ms ~647 ms
#201	Qwen3-14B	14B	INT4 Q8	78 tok/s 48 tok/s	~1.1s ~1.1s
#203	Command R	35B	INT4	26 tok/s	~3.1s
#205	Llama 3 8B	8B	INT4 Q8 FP16	117 tok/s 76 tok/s 43 tok/s	~633 ms ~637 ms ~647 ms
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	326 tok/s 311 tok/s 282 tok/s	~30 ms ~30 ms ~31 ms
#207	Qwen3-1.7B	1.7B	INT4 Q8 FP16	246 tok/s 199 tok/s 140 tok/s	~153 ms ~154 ms ~156 ms
#208	Qwen3-0.6B	0.6B	INT4 Q8 FP16	315 tok/s 285 tok/s 234 tok/s	~58 ms ~59 ms ~59 ms
#209	Phi-4-Mini	3.8B	INT4 Q8 FP16	180 tok/s 129 tok/s 80 tok/s	~314 ms ~316 ms ~321 ms
#210	Qwen3-8B	8B	INT4 Q8 FP16	117 tok/s 76 tok/s 43 tok/s	~633 ms ~637 ms ~647 ms
#212	Ministral-8B-2410	8B	INT4 Q8 FP16	117 tok/s 76 tok/s 43 tok/s	~633 ms ~637 ms ~647 ms
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	256 tok/s 210 tok/s 151 tok/s	~136 ms ~137 ms ~139 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	200 tok/s 149 tok/s 95 tok/s	~253 ms ~255 ms ~259 ms
#215	Phi-3-mini	3.8B	INT4 Q8 FP16	180 tok/s 129 tok/s 80 tok/s	~314 ms ~316 ms ~321 ms
#216	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	299 tok/s 264 tok/s 208 tok/s	~76 ms ~77 ms ~78 ms
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	148 tok/s 130 tok/s 98 tok/s	~192 ms ~193 ms ~195 ms
#218	Gemma 2 2B	2B	INT4 Q8 FP16	233 tok/s 184 tok/s 126 tok/s	~178 ms ~179 ms ~181 ms
#219	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	124 tok/s 82 tok/s 47 tok/s	~579 ms ~583 ms ~593 ms
#220	Llama 3.2 1B	1B	INT4 Q8 FP16	284 tok/s 245 tok/s 187 tok/s	~95 ms ~96 ms ~97 ms
#221	Phi-3-medium	14B	INT4 Q8	78 tok/s 48 tok/s	~1.1s ~1.1s
#222	Yi-34B	34B	INT4	26 tok/s	~3.1s
#223	OLMo 3 7B Instruct	7B	INT4 Q8 FP16	96 tok/s 70 tok/s 43 tok/s	~559 ms ~563 ms ~572 ms
#224	Gemma 1 7B	7B	INT4 Q8 FP16	96 tok/s 70 tok/s 43 tok/s	~559 ms ~563 ms ~572 ms
#225	Qwen2-7B	7B	INT4 Q8 FP16	127 tok/s 85 tok/s 48 tok/s	~557 ms ~561 ms ~570 ms
#226	Phi-3-small	7B	INT4 Q8 FP16	127 tok/s 85 tok/s 48 tok/s	~557 ms ~561 ms ~570 ms
#227	Gemma 3 1B	1B	INT4 Q8 FP16	284 tok/s 245 tok/s 187 tok/s	~95 ms ~96 ms ~97 ms
#228	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	124 tok/s 82 tok/s 47 tok/s	~579 ms ~583 ms ~593 ms
#230	Gemma 1 2B	2B	INT4 Q8 FP16	243 tok/s 190 tok/s 129 tok/s	~178 ms ~179 ms ~181 ms
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	105 tok/s 79 tok/s 49 tok/s	~483 ms ~486 ms ~494 ms
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	105 tok/s 79 tok/s 49 tok/s	~483 ms ~486 ms ~494 ms
#233	ChatGLM-6B	6B	INT4 Q8 FP16	105 tok/s 79 tok/s 49 tok/s	~483 ms ~486 ms ~494 ms
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	323 tok/s 296 tok/s 250 tok/s	~49 ms ~50 ms ~50 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	323 tok/s 296 tok/s 250 tok/s	~49 ms ~50 ms ~50 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	320 tok/s 305 tok/s 274 tok/s	~33 ms ~33 ms ~33 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	320 tok/s 305 tok/s 274 tok/s	~33 ms ~33 ms ~33 ms
-	Falcon-1B	1B	INT4 Q8 FP16	293 tok/s 251 tok/s 191 tok/s	~95 ms ~96 ms ~97 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	284 tok/s 245 tok/s 187 tok/s	~95 ms ~96 ms ~97 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	256 tok/s 210 tok/s 151 tok/s	~136 ms ~137 ms ~139 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	256 tok/s 210 tok/s 151 tok/s	~136 ms ~137 ms ~139 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	246 tok/s 199 tok/s 140 tok/s	~153 ms ~154 ms ~156 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	242 tok/s 194 tok/s 135 tok/s	~161 ms ~162 ms ~164 ms
-	Falcon-3B	3B	INT4 Q8 FP16	208 tok/s 153 tok/s 97 tok/s	~253 ms ~255 ms ~258 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	205 tok/s 181 tok/s 140 tok/s	~121 ms ~121 ms ~123 ms
-	Phi-1	1.3B	INT4 Q8 FP16	205 tok/s 181 tok/s 140 tok/s	~121 ms ~121 ms ~123 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	205 tok/s 181 tok/s 140 tok/s	~121 ms ~121 ms ~123 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	202 tok/s 150 tok/s 96 tok/s	~253 ms ~255 ms ~258 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	200 tok/s 149 tok/s 95 tok/s	~253 ms ~255 ms ~259 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	200 tok/s 149 tok/s 95 tok/s	~253 ms ~255 ms ~259 ms
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4 Q8	181 tok/s 125 tok/s	~308 ms ~363 ms
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4 Q8	181 tok/s 125 tok/s	~308 ms ~363 ms
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4	177 tok/s	~323 ms
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4	177 tok/s	~323 ms
-	Hy-MT2 30B A3B	30B (3B active)	INT4	175 tok/s	~328 ms
-	Spark X2.5 4B	4B	INT4 Q8 FP16	175 tok/s 125 tok/s 77 tok/s	~329 ms ~331 ms ~336 ms
-	Phi-2	2.7B	INT4 Q8 FP16	154 tok/s 127 tok/s 89 tok/s	~232 ms ~234 ms ~237 ms
-	MaLLaM-3B	3B	INT4 Q8 FP16	148 tok/s 120 tok/s 83 tok/s	~255 ms ~256 ms ~260 ms
-	SmolLM3 3B	3B	INT4 Q8 FP16	148 tok/s 120 tok/s 83 tok/s	~255 ms ~256 ms ~260 ms
-	Falcon-7B	7B	INT4 Q8 FP16	132 tok/s 86 tok/s 49 tok/s	~557 ms ~561 ms ~569 ms
-	Falcon3-7B	7B	INT4 Q8 FP16	127 tok/s 85 tok/s 48 tok/s	~557 ms ~561 ms ~570 ms
-	Hy-MT2 7B	7B	INT4 Q8 FP16	127 tok/s 85 tok/s 48 tok/s	~557 ms ~561 ms ~570 ms
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	124 tok/s 82 tok/s 47 tok/s	~579 ms ~583 ms ~593 ms
-	LensVLM 9B	9B	INT4 Q8 FP16	108 tok/s 69 tok/s 37 tok/s	~710 ms ~715 ms ~788 ms
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8 FP16	108 tok/s 69 tok/s 37 tok/s	~710 ms ~715 ms ~788 ms
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	105 tok/s 79 tok/s 49 tok/s	~483 ms ~486 ms ~494 ms
-	Yi-6B	6B	INT4 Q8 FP16	105 tok/s 79 tok/s 49 tok/s	~483 ms ~486 ms ~494 ms
-	Falcon3-10B	10B	INT4 Q8 FP16	100 tok/s 64 tok/s 34 tok/s	~786 ms ~792 ms ~872 ms
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4 Q8	100 tok/s 84 tok/s	~300 ms ~327 ms
-	Falcon2-11B	11B	INT4 Q8 FP16	97 tok/s 60 tok/s 30 tok/s	~862 ms ~868 ms ~1.0s
-	Hunyuan Lite	7B	INT4 Q8 FP16	96 tok/s 70 tok/s 43 tok/s	~559 ms ~563 ms ~572 ms
-	OLMo 3 7B Base	7B	INT4 Q8 FP16	96 tok/s 70 tok/s 43 tok/s	~559 ms ~563 ms ~572 ms
-	SEA-LION-7B	7.1B	INT4 Q8 FP16	95 tok/s 70 tok/s 43 tok/s	~567 ms ~571 ms ~580 ms
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8 FP16	95 tok/s 70 tok/s 43 tok/s	~567 ms ~571 ms ~580 ms
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8 FP16	88 tok/s 64 tok/s 37 tok/s	~635 ms ~640 ms ~705 ms
-	Typhoon-2-8B	8B	INT4 Q8 FP16	88 tok/s 64 tok/s 37 tok/s	~635 ms ~640 ms ~705 ms
-	GLM-4-9B	9B	INT4 Q8 FP16	81 tok/s 58 tok/s 33 tok/s	~713 ms ~718 ms ~791 ms
-	GLM-4-9B-Chat	9B	INT4 Q8 FP16	81 tok/s 58 tok/s 33 tok/s	~713 ms ~718 ms ~791 ms
-	GLM-4-9B-Chat-1M	9B	INT4 Q8 FP16	81 tok/s 58 tok/s 33 tok/s	~713 ms ~718 ms ~791 ms
-	Yi-9B	9B	INT4 Q8 FP16	81 tok/s 58 tok/s 33 tok/s	~713 ms ~718 ms ~791 ms
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8 FP16	80 tok/s 57 tok/s 33 tok/s	~728 ms ~733 ms ~807 ms
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8 FP16	80 tok/s 57 tok/s 33 tok/s	~728 ms ~733 ms ~807 ms
-	GreenMind-14B-R1	14B	INT4 Q8	59 tok/s 40 tok/s	~1.1s ~1.1s
-	Falcon-40B	40B	INT4	31 tok/s	~3.3s

NVIDIA RTX 3090 (24GB)

VRAM Scaling by Context Length

Best Models to Run on NVIDIA RTX 3090 (24GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions