What are the best models to run on NVIDIA RTX 3070 (8GB)?

Based on model rankings and 8 GB usable VRAM: Top Overall: Qwen3.5-9B (INT4) ~61 tok/s. Top Coding: K2 Horizon 7B (INT4) ~74 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~67 tok/s. Maximum Model Size: Mellum 2.1 12B A2.5B Thinking (INT4) ~122 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 LLMs to Run on NVIDIA RTX 3070 (8GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#121	Qwen3.5-9B	9B	INT4	61 tok/s	~1.2s
#127	K2 Horizon 7B	7B	INT4	74 tok/s	~972 ms
#136	Qwen3.5-4B	4B	INT4 Q8	111 tok/s 73 tok/s	~572 ms ~577 ms
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	20 tok/s 11 tok/s	~3.9s ~4.3s
#152	Granite 4.2 8B	8B	INT4	67 tok/s	~1.1s
#157	Gemma 4 E4B	8B	INT4	67 tok/s	~1.1s
#180	Gemma 3 12B	12B	INT4	43 tok/s	~1.9s
#185	Ministral 3 8B	8B	INT4	43 tok/s	~1.3s
#187	OLMo 3 7B Think	7B	INT4	50 tok/s	~1.1s
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	166 tok/s 119 tok/s 73 tok/s	~307 ms ~309 ms ~314 ms
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	166 tok/s 119 tok/s 73 tok/s	~307 ms ~309 ms ~314 ms
#193	Gemma 4 E2B	5.1B	INT4 Q8	94 tok/s 57 tok/s	~719 ms ~787 ms
#195	Ministral 3 3B	3B	INT4 Q8	89 tok/s 69 tok/s	~443 ms ~446 ms
#196	DeepSeek-R1 8B	8B	INT4	69 tok/s	~1.1s
#198	Gemma 2 9B	9B	INT4	61 tok/s	~1.2s
#201	Qwen3-4B	4B	INT4 Q8	111 tok/s 73 tok/s	~572 ms ~577 ms
#203	Gemma 3 4B	4B	INT4 Q8	111 tok/s 73 tok/s	~572 ms ~577 ms
#204	DeepSeek-R1 7B	7B	INT4 Q8	76 tok/s 42 tok/s	~972 ms ~1.2s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	247 tok/s 200 tok/s 141 tok/s	~128 ms ~129 ms ~131 ms
#207	Llama 3.1 8B	8B	INT4	67 tok/s	~1.1s
#207	Qwen3-8B	8B	INT4	67 tok/s	~1.1s
#212	Llama 3 8B	8B	INT4	67 tok/s	~1.1s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	180 tok/s 132 tok/s 83 tok/s	~263 ms ~265 ms ~270 ms
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	288 tok/s 265 tok/s 223 tok/s	~47 ms ~47 ms ~48 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	271 tok/s 227 tok/s 167 tok/s	~97 ms ~97 ms ~99 ms
#217	Phi-4-Mini	3.8B	INT4 Q8	115 tok/s 76 tok/s	~546 ms ~550 ms
#218	Qwen2.5-7B	7B	INT4	74 tok/s	~972 ms
#220	Ministral-8B-2410	8B	INT4	67 tok/s	~1.1s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	192 tok/s 143 tok/s 92 tok/s	~233 ms ~235 ms ~239 ms
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
#223	Phi-3-mini	3.8B	INT4 Q8	115 tok/s 76 tok/s	~546 ms ~550 ms
#224	Gemma 3n E2B IT	6B (2B active)	INT4	85 tok/s	~360 ms
#225	Gemma 2 2B	2B	INT4 Q8 FP16	166 tok/s 119 tok/s 73 tok/s	~307 ms ~309 ms ~314 ms
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4	72 tok/s	~1.0s
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	227 tok/s 179 tok/s 122 tok/s	~161 ms ~163 ms ~165 ms
#230	OLMo 3 7B Instruct	7B	INT4	50 tok/s	~1.1s
#231	Qwen2-7B	7B	INT4	74 tok/s	~972 ms
#232	Phi-3-small	7B	INT4	74 tok/s	~972 ms
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4	72 tok/s	~1.0s
#234	Gemma 3 1B	1B	INT4 Q8 FP16	227 tok/s 179 tok/s 122 tok/s	~161 ms ~163 ms ~165 ms
#236	Gemma 1 7B	7B	INT4	50 tok/s	~1.1s
#237	Gemma 1 2B	2B	INT4 Q8 FP16	176 tok/s 124 tok/s 75 tok/s	~306 ms ~309 ms ~314 ms
#238	ChatGLM3-6B	6B	INT4	56 tok/s	~916 ms
#239	ChatGLM2-6B	6B	INT4	56 tok/s	~916 ms
#240	ChatGLM-6B	6B	INT4	56 tok/s	~916 ms
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	314 tok/s 282 tok/s 231 tok/s	~51 ms ~52 ms ~52 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	284 tok/s 243 tok/s 184 tok/s	~81 ms ~82 ms ~83 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	284 tok/s 243 tok/s 184 tok/s	~81 ms ~82 ms ~83 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	279 tok/s 256 tok/s 213 tok/s	~52 ms ~52 ms ~53 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	279 tok/s 256 tok/s 213 tok/s	~52 ms ~52 ms ~53 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	271 tok/s 227 tok/s 167 tok/s	~97 ms ~97 ms ~99 ms
-	Falcon-1B	1B	INT4 Q8 FP16	238 tok/s 186 tok/s 125 tok/s	~161 ms ~162 ms ~165 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	227 tok/s 179 tok/s 122 tok/s	~161 ms ~163 ms ~165 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	227 tok/s 179 tok/s 122 tok/s	~161 ms ~163 ms ~165 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	192 tok/s 143 tok/s 92 tok/s	~233 ms ~235 ms ~239 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	192 tok/s 143 tok/s 92 tok/s	~233 ms ~235 ms ~239 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	180 tok/s 132 tok/s 83 tok/s	~263 ms ~265 ms ~270 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	175 tok/s 128 tok/s 80 tok/s	~277 ms ~279 ms ~284 ms
-	Falcon-3B	3B	INT4 Q8 FP16	141 tok/s 94 tok/s 51 tok/s	~438 ms ~442 ms ~488 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	138 tok/s 117 tok/s 83 tok/s	~206 ms ~208 ms ~211 ms
-	Phi-1	1.3B	INT4 Q8 FP16	138 tok/s 117 tok/s 83 tok/s	~206 ms ~208 ms ~211 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	138 tok/s 117 tok/s 83 tok/s	~206 ms ~208 ms ~211 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	135 tok/s 91 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	LFM2-VL 3B	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	Liquid D1 3B	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	133 tok/s 90 tok/s 50 tok/s	~439 ms ~442 ms ~488 ms
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	122 tok/s	~503 ms
-	FrogNano 4B	4B	INT4 Q8	111 tok/s 73 tok/s	~572 ms ~577 ms
-	Spark X2.5 4B	4B	INT4 Q8	111 tok/s 73 tok/s	~572 ms ~577 ms
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	107 tok/s 87 tok/s 57 tok/s	~310 ms ~312 ms ~344 ms
-	Phi-2	2.7B	INT4 Q8 FP16	94 tok/s 74 tok/s 44 tok/s	~403 ms ~406 ms ~483 ms
-	MaLLaM-3B	3B	INT4 Q8	89 tok/s 69 tok/s	~443 ms ~446 ms
-	SmolLM3 3B	3B	INT4 Q8	89 tok/s 69 tok/s	~443 ms ~446 ms
-	Falcon-7B	7B	INT4 Q8	78 tok/s 42 tok/s	~972 ms ~1.1s
-	Falcon3-7B	7B	INT4	74 tok/s	~972 ms
-	Hy-MT2 7B	7B	INT4	74 tok/s	~972 ms
-	Mistral-7B-v0.1	7.3B	INT4	72 tok/s	~1.0s
-	Youtu Parsing Omni	8B	INT4	69 tok/s	~1.1s
-	LensVLM 9B	9B	INT4	61 tok/s	~1.2s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	61 tok/s	~1.2s
-	ChatGLM3-6B-32K	6B	INT4	56 tok/s	~916 ms
-	Yi-6B	6B	INT4	56 tok/s	~916 ms
-	Falcon3-10B	10B	INT4	53 tok/s	~1.5s
-	Falcon2-11B	11B	INT4	51 tok/s	~1.6s
-	Hunyuan Lite	7B	INT4	50 tok/s	~1.1s
-	OLMo 3 7B Base	7B	INT4	50 tok/s	~1.1s
-	SEA-LION-7B	7.1B	INT4	50 tok/s	~1.1s
-	SEA-LION-7B-Instruct	7.1B	INT4	50 tok/s	~1.1s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	43 tok/s	~1.3s
-	Typhoon-2-8B	8B	INT4	43 tok/s	~1.3s

NVIDIA RTX 3070 (8GB)

VRAM Scaling by Context Length

Best Models to Run on NVIDIA RTX 3070 (8GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions