What are the best models to run on NVIDIA RTX 5070 (12GB)?

Based on model rankings and 12 GB usable VRAM: Top Overall: Phi-4 Reasoning Plus (INT4) ~40 tok/s. Top Coding: K2 Horizon 7B (INT4) ~101 tok/s. Optimal Daily Driver: Gemma 4 12B (INT4) ~47 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 5070 (12GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#100	Phi-4 Reasoning Plus	14B	INT4	40 tok/s	~1.5s
#109	Gemma 4 12B	11.95B	INT4	47 tok/s	~1.2s
#113	Qwen3.5-9B	9B	INT4 Q8	84 tok/s 50 tok/s	~820 ms ~898 ms
#118	K2 Horizon 7B	7B	INT4 Q8	101 tok/s 65 tok/s	~643 ms ~649 ms
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	29 tok/s 17 tok/s 9 tok/s	~2.6s ~2.6s ~2.9s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	145 tok/s 99 tok/s 59 tok/s	~380 ms ~383 ms ~390 ms
#146	Granite 4.2 8B	8B	INT4 Q8	92 tok/s 55 tok/s	~731 ms ~800 ms
#150	Gemma 4 E4B	8B	INT4 Q8	92 tok/s 55 tok/s	~731 ms ~800 ms
#158	GLM-4V	9B	INT4	62 tok/s	~825 ms
#160	Phi-4	14B	INT4	59 tok/s	~1.3s
#169	OLMo 3 7B Think	7B	INT4 Q8	74 tok/s 51 tok/s	~647 ms ~707 ms
#172	DeepSeek-R1 14B	14B	INT4	61 tok/s	~1.3s
#173	Ministral 3 14B	14B	INT4	40 tok/s	~1.5s
#177	Gemma 3 12B	12B	INT4	67 tok/s	~1.1s
#182	Ministral 3 8B	8B	INT4 Q8	68 tok/s 43 tok/s	~735 ms ~869 ms
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	203 tok/s 154 tok/s 100 tok/s	~205 ms ~206 ms ~210 ms
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	203 tok/s 154 tok/s 100 tok/s	~205 ms ~206 ms ~210 ms
#185	Qwen2.5-14B	14B	INT4	59 tok/s	~1.3s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	125 tok/s 83 tok/s 43 tok/s	~476 ms ~480 ms ~573 ms
#189	Qwen3-4B	4B	INT4 Q8 FP16	145 tok/s 99 tok/s 59 tok/s	~380 ms ~383 ms ~390 ms
#190	DeepSeek-R1 8B	8B	INT4 Q8	94 tok/s 59 tok/s	~731 ms ~737 ms
#193	Ministral 3 3B	3B	INT4 Q8 FP16	120 tok/s 95 tok/s 63 tok/s	~294 ms ~296 ms ~302 ms
#195	Qwen3-14B	14B	INT4	59 tok/s	~1.3s
#197	Gemma 2 9B	9B	INT4 Q8	84 tok/s 50 tok/s	~820 ms ~898 ms
#198	Qwen2.5-7B	7B	INT4 Q8	101 tok/s 65 tok/s	~643 ms ~649 ms
#199	Qwen3-8B	8B	INT4 Q8	92 tok/s 55 tok/s	~731 ms ~800 ms
#200	Gemma 3 4B	4B	INT4 Q8 FP16	145 tok/s 99 tok/s 59 tok/s	~380 ms ~383 ms ~390 ms
#201	DeepSeek-R1 7B	7B	INT4 Q8	104 tok/s 66 tok/s	~643 ms ~649 ms
#202	Llama 3.1 8B	8B	INT4 Q8	92 tok/s 55 tok/s	~731 ms ~800 ms
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	217 tok/s 168 tok/s 112 tok/s	~176 ms ~177 ms ~180 ms
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	311 tok/s 293 tok/s 257 tok/s	~34 ms ~34 ms ~34 ms
#208	Llama 3 8B	8B	INT4 Q8	92 tok/s 55 tok/s	~731 ms ~800 ms
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	298 tok/s 261 tok/s 205 tok/s	~66 ms ~67 ms ~68 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	149 tok/s 103 tok/s 61 tok/s	~362 ms ~365 ms ~372 ms
#212	Ministral-8B-2410	8B	INT4 Q8	92 tok/s 55 tok/s	~731 ms ~800 ms
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	228 tok/s 180 tok/s 122 tok/s	~156 ms ~158 ms ~160 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	169 tok/s 121 tok/s 74 tok/s	~292 ms ~294 ms ~299 ms
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	278 tok/s 237 tok/s 178 tok/s	~87 ms ~88 ms ~89 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	149 tok/s 103 tok/s 61 tok/s	~362 ms ~365 ms ~372 ms
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8	119 tok/s 103 tok/s	~221 ms ~223 ms
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8	98 tok/s 63 tok/s	~669 ms ~675 ms
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	260 tok/s 216 tok/s 157 tok/s	~109 ms ~110 ms ~112 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	203 tok/s 154 tok/s 100 tok/s	~205 ms ~206 ms ~210 ms
#221	Phi-3-medium	14B	INT4	59 tok/s	~1.3s
#222	OLMo 3 7B Instruct	7B	INT4 Q8	74 tok/s 51 tok/s	~647 ms ~707 ms
#224	Qwen2-7B	7B	INT4 Q8	101 tok/s 65 tok/s	~643 ms ~649 ms
#225	Phi-3-small	7B	INT4 Q8	101 tok/s 65 tok/s	~643 ms ~649 ms
#226	Gemma 3 1B	1B	INT4 Q8 FP16	260 tok/s 216 tok/s 157 tok/s	~109 ms ~110 ms ~112 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8	98 tok/s 63 tok/s	~669 ms ~675 ms
#229	Gemma 1 7B	7B	INT4 Q8	74 tok/s 51 tok/s	~647 ms ~707 ms
#230	Gemma 1 2B	2B	INT4 Q8 FP16	214 tok/s 159 tok/s 102 tok/s	~204 ms ~206 ms ~209 ms
#231	ChatGLM3-6B	6B	INT4 Q8	82 tok/s 60 tok/s	~559 ms ~563 ms
#232	ChatGLM2-6B	6B	INT4 Q8	82 tok/s 60 tok/s	~559 ms ~563 ms
#233	ChatGLM-6B	6B	INT4 Q8	82 tok/s 60 tok/s	~559 ms ~563 ms
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	308 tok/s 275 tok/s 221 tok/s	~56 ms ~56 ms ~57 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	308 tok/s 275 tok/s 221 tok/s	~56 ms ~56 ms ~57 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	304 tok/s 285 tok/s 248 tok/s	~37 ms ~37 ms ~37 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	304 tok/s 285 tok/s 248 tok/s	~37 ms ~37 ms ~37 ms
-	Falcon-1B	1B	INT4 Q8 FP16	270 tok/s 223 tok/s 160 tok/s	~109 ms ~110 ms ~111 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	260 tok/s 216 tok/s 157 tok/s	~109 ms ~110 ms ~112 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	228 tok/s 180 tok/s 122 tok/s	~156 ms ~158 ms ~160 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	228 tok/s 180 tok/s 122 tok/s	~156 ms ~158 ms ~160 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	217 tok/s 168 tok/s 112 tok/s	~176 ms ~177 ms ~180 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	213 tok/s 164 tok/s 108 tok/s	~185 ms ~186 ms ~189 ms
-	Falcon-3B	3B	INT4 Q8 FP16	177 tok/s 125 tok/s 75 tok/s	~291 ms ~294 ms ~299 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	175 tok/s 151 tok/s 112 tok/s	~139 ms ~139 ms ~142 ms
-	Phi-1	1.3B	INT4 Q8 FP16	175 tok/s 151 tok/s 112 tok/s	~139 ms ~139 ms ~142 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	175 tok/s 151 tok/s 112 tok/s	~139 ms ~139 ms ~142 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	171 tok/s 122 tok/s 74 tok/s	~292 ms ~294 ms ~299 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	169 tok/s 121 tok/s 74 tok/s	~292 ms ~294 ms ~299 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	169 tok/s 121 tok/s 74 tok/s	~292 ms ~294 ms ~299 ms
-	Spark X2.5 4B	4B	INT4 Q8 FP16	145 tok/s 99 tok/s 59 tok/s	~380 ms ~383 ms ~390 ms
-	Phi-2	2.7B	INT4 Q8 FP16	125 tok/s 101 tok/s 68 tok/s	~268 ms ~270 ms ~275 ms
-	MaLLaM-3B	3B	INT4 Q8 FP16	120 tok/s 95 tok/s 63 tok/s	~294 ms ~296 ms ~302 ms
-	SmolLM3 3B	3B	INT4 Q8 FP16	120 tok/s 95 tok/s 63 tok/s	~294 ms ~296 ms ~302 ms
-	Falcon-7B	7B	INT4 Q8	105 tok/s 66 tok/s	~643 ms ~649 ms
-	Falcon3-7B	7B	INT4 Q8	101 tok/s 65 tok/s	~643 ms ~649 ms
-	Hy-MT2 7B	7B	INT4 Q8	101 tok/s 65 tok/s	~643 ms ~649 ms
-	Mistral-7B-v0.1	7.3B	INT4 Q8	98 tok/s 63 tok/s	~669 ms ~675 ms
-	LensVLM 9B	9B	INT4 Q8	84 tok/s 50 tok/s	~820 ms ~898 ms
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	84 tok/s 50 tok/s	~820 ms ~898 ms
-	ChatGLM3-6B-32K	6B	INT4 Q8	82 tok/s 60 tok/s	~559 ms ~563 ms
-	Yi-6B	6B	INT4 Q8	82 tok/s 60 tok/s	~559 ms ~563 ms
-	Falcon3-10B	10B	INT4 Q8	78 tok/s 43 tok/s	~908 ms ~1.1s
-	Falcon2-11B	11B	INT4 Q8	75 tok/s 41 tok/s	~996 ms ~1.2s
-	Hunyuan Lite	7B	INT4 Q8	74 tok/s 51 tok/s	~647 ms ~707 ms
-	OLMo 3 7B Base	7B	INT4 Q8	74 tok/s 51 tok/s	~647 ms ~707 ms
-	SEA-LION-7B	7.1B	INT4 Q8	74 tok/s 50 tok/s	~656 ms ~717 ms
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	74 tok/s 50 tok/s	~656 ms ~717 ms
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	68 tok/s 43 tok/s	~735 ms ~869 ms
-	Typhoon-2-8B	8B	INT4 Q8	68 tok/s 43 tok/s	~735 ms ~869 ms
-	GLM-4-9B	9B	INT4	62 tok/s	~825 ms
-	GLM-4-9B-Chat	9B	INT4	62 tok/s	~825 ms
-	GLM-4-9B-Chat-1M	9B	INT4	62 tok/s	~825 ms
-	Yi-9B	9B	INT4	62 tok/s	~825 ms
-	Sahabat-AI-Gemma2-9B	9.2B	INT4	61 tok/s	~842 ms
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4	61 tok/s	~842 ms
-	GreenMind-14B-R1	14B	INT4	40 tok/s	~1.5s

NVIDIA RTX 5070 (12GB)

VRAM Scaling by Context Length

Best Models to Run on NVIDIA RTX 5070 (12GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions