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

Based on model rankings and 8 GB usable VRAM: Top Overall: Qwen3.5-9B (INT4) ~78 tok/s. Top Coding: K2 Horizon 7B (INT4) ~94 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~85 tok/s. Maximum Model Size: Mellum 2.1 12B A2.5B Thinking (INT4) ~146 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 Ti (8GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#121	Qwen3.5-9B	9B	INT4	78 tok/s	~1.2s
#127	K2 Horizon 7B	7B	INT4	94 tok/s	~906 ms
#136	Qwen3.5-4B	4B	INT4 Q8	136 tok/s 92 tok/s	~533 ms ~537 ms
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	27 tok/s 15 tok/s	~3.6s ~4.0s
#152	Granite 4.2 8B	8B	INT4	85 tok/s	~1.0s
#157	Gemma 4 E4B	8B	INT4	85 tok/s	~1.0s
#180	Gemma 3 12B	12B	INT4	56 tok/s	~1.8s
#185	Ministral 3 8B	8B	INT4	56 tok/s	~1.2s
#187	OLMo 3 7B Think	7B	INT4	65 tok/s	~988 ms
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	194 tok/s 145 tok/s 93 tok/s	~286 ms ~288 ms ~292 ms
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	194 tok/s 145 tok/s 93 tok/s	~286 ms ~288 ms ~292 ms
#193	Gemma 4 E2B	5.1B	INT4 Q8	117 tok/s 73 tok/s	~670 ms ~732 ms
#195	Ministral 3 3B	3B	INT4 Q8	112 tok/s 88 tok/s	~412 ms ~414 ms
#196	DeepSeek-R1 8B	8B	INT4	87 tok/s	~1.0s
#198	Gemma 2 9B	9B	INT4	78 tok/s	~1.2s
#201	Qwen3-4B	4B	INT4 Q8	136 tok/s 92 tok/s	~533 ms ~537 ms
#203	Gemma 3 4B	4B	INT4 Q8	136 tok/s 92 tok/s	~533 ms ~537 ms
#204	DeepSeek-R1 7B	7B	INT4 Q8	96 tok/s 54 tok/s	~906 ms ~1.1s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	271 tok/s 228 tok/s 169 tok/s	~120 ms ~121 ms ~122 ms
#207	Llama 3.1 8B	8B	INT4	85 tok/s	~1.0s
#207	Qwen3-8B	8B	INT4	85 tok/s	~1.0s
#212	Llama 3 8B	8B	INT4	85 tok/s	~1.0s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	208 tok/s 159 tok/s 105 tok/s	~246 ms ~247 ms ~250 ms
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	306 tok/s 287 tok/s 249 tok/s	~44 ms ~44 ms ~45 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	292 tok/s 253 tok/s 196 tok/s	~90 ms ~91 ms ~92 ms
#217	Phi-4-Mini	3.8B	INT4 Q8	140 tok/s 96 tok/s	~509 ms ~512 ms
#218	Qwen2.5-7B	7B	INT4	94 tok/s	~906 ms
#220	Ministral-8B-2410	8B	INT4	85 tok/s	~1.0s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	220 tok/s 171 tok/s 114 tok/s	~218 ms ~219 ms ~222 ms
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
#223	Phi-3-mini	3.8B	INT4 Q8	140 tok/s 96 tok/s	~509 ms ~512 ms
#224	Gemma 3n E2B IT	6B (2B active)	INT4	106 tok/s	~335 ms
#225	Gemma 2 2B	2B	INT4 Q8 FP16	194 tok/s 145 tok/s 93 tok/s	~286 ms ~288 ms ~292 ms
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4	91 tok/s	~943 ms
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	252 tok/s 207 tok/s 148 tok/s	~151 ms ~152 ms ~154 ms
#230	OLMo 3 7B Instruct	7B	INT4	65 tok/s	~988 ms
#231	Qwen2-7B	7B	INT4	94 tok/s	~906 ms
#232	Phi-3-small	7B	INT4	94 tok/s	~906 ms
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4	91 tok/s	~943 ms
#234	Gemma 3 1B	1B	INT4 Q8 FP16	252 tok/s 207 tok/s 148 tok/s	~151 ms ~152 ms ~154 ms
#236	Gemma 1 7B	7B	INT4	65 tok/s	~988 ms
#237	Gemma 1 2B	2B	INT4 Q8 FP16	205 tok/s 150 tok/s 95 tok/s	~286 ms ~287 ms ~291 ms
#238	ChatGLM3-6B	6B	INT4	72 tok/s	~853 ms
#239	ChatGLM2-6B	6B	INT4	72 tok/s	~853 ms
#240	ChatGLM-6B	6B	INT4	72 tok/s	~853 ms
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	327 tok/s 301 tok/s 256 tok/s	~48 ms ~48 ms ~49 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	303 tok/s 267 tok/s 212 tok/s	~76 ms ~76 ms ~77 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	303 tok/s 267 tok/s 212 tok/s	~76 ms ~76 ms ~77 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	298 tok/s 279 tok/s 240 tok/s	~49 ms ~49 ms ~49 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	298 tok/s 279 tok/s 240 tok/s	~49 ms ~49 ms ~49 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	292 tok/s 253 tok/s 196 tok/s	~90 ms ~91 ms ~92 ms
-	Falcon-1B	1B	INT4 Q8 FP16	263 tok/s 214 tok/s 151 tok/s	~151 ms ~152 ms ~153 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	252 tok/s 207 tok/s 148 tok/s	~151 ms ~152 ms ~154 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	252 tok/s 207 tok/s 148 tok/s	~151 ms ~152 ms ~154 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	220 tok/s 171 tok/s 114 tok/s	~218 ms ~219 ms ~222 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	220 tok/s 171 tok/s 114 tok/s	~218 ms ~219 ms ~222 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	208 tok/s 159 tok/s 105 tok/s	~246 ms ~247 ms ~250 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	204 tok/s 154 tok/s 100 tok/s	~258 ms ~260 ms ~263 ms
-	Falcon-3B	3B	INT4 Q8 FP16	168 tok/s 116 tok/s 66 tok/s	~409 ms ~412 ms ~453 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	166 tok/s 142 tok/s 105 tok/s	~192 ms ~193 ms ~196 ms
-	Phi-1	1.3B	INT4 Q8 FP16	166 tok/s 142 tok/s 105 tok/s	~192 ms ~193 ms ~196 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	166 tok/s 142 tok/s 105 tok/s	~192 ms ~193 ms ~196 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	162 tok/s 114 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	LFM2-VL 3B	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	Liquid D1 3B	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	160 tok/s 113 tok/s 65 tok/s	~409 ms ~412 ms ~453 ms
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	146 tok/s	~469 ms
-	FrogNano 4B	4B	INT4 Q8	136 tok/s 92 tok/s	~533 ms ~537 ms
-	Spark X2.5 4B	4B	INT4 Q8	136 tok/s 92 tok/s	~533 ms ~537 ms
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	131 tok/s 109 tok/s 73 tok/s	~288 ms ~290 ms ~319 ms
-	Phi-2	2.7B	INT4 Q8 FP16	117 tok/s 93 tok/s 57 tok/s	~375 ms ~377 ms ~448 ms
-	MaLLaM-3B	3B	INT4 Q8	112 tok/s 88 tok/s	~412 ms ~414 ms
-	SmolLM3 3B	3B	INT4 Q8	112 tok/s 88 tok/s	~412 ms ~414 ms
-	Falcon-7B	7B	INT4 Q8	98 tok/s 55 tok/s	~906 ms ~1.1s
-	Falcon3-7B	7B	INT4	94 tok/s	~906 ms
-	Hy-MT2 7B	7B	INT4	94 tok/s	~906 ms
-	Mistral-7B-v0.1	7.3B	INT4	91 tok/s	~943 ms
-	Youtu Parsing Omni	8B	INT4	87 tok/s	~1.0s
-	LensVLM 9B	9B	INT4	78 tok/s	~1.2s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	78 tok/s	~1.2s
-	ChatGLM3-6B-32K	6B	INT4	72 tok/s	~853 ms
-	Yi-6B	6B	INT4	72 tok/s	~853 ms
-	Falcon3-10B	10B	INT4	68 tok/s	~1.4s
-	Falcon2-11B	11B	INT4	66 tok/s	~1.5s
-	Hunyuan Lite	7B	INT4	65 tok/s	~988 ms
-	OLMo 3 7B Base	7B	INT4	65 tok/s	~988 ms
-	SEA-LION-7B	7.1B	INT4	65 tok/s	~1.0s
-	SEA-LION-7B-Instruct	7.1B	INT4	65 tok/s	~1.0s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	56 tok/s	~1.2s
-	Typhoon-2-8B	8B	INT4	56 tok/s	~1.2s

NVIDIA RTX 3070 Ti (8GB)

VRAM Scaling by Context Length

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

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions