What are the best models to run on NVIDIA A2 (16GB)?

Based on model rankings and 16 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~9 tok/s. Top Coding: MiMo V2 Flash (INT4) ~1 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~14 tok/s. Maximum Model Size: ERNIE-4.5-VL-28B-A3B (INT4) ~55 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 A2 (16GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#36	Qwen 3.8 27B	27B	INT4	9 tok/s	~21.8s
#86	Qwen3.5-27B	27B	INT4	9 tok/s	~21.8s
#102	Phi-4 Reasoning Plus	14B	INT4	14 tok/s	~9.8s
#105	MiMo V2 Flash	15B (309B active)	INT4	1 tok/s	~220.2s
#110	Gemma 4 26B A4B	25.2B (3.8B active)	INT4	51 tok/s	~3.6s
#121	Qwen3.5-9B	9B	INT4 Q8	29 tok/s 17 tok/s	~6.3s ~6.3s
#127	K2 Horizon 7B	7B	INT4 Q8 FP16	37 tok/s 21 tok/s 10 tok/s	~4.9s ~5.0s ~5.9s
#130	Gemma 4 12B	11.95B	INT4 Q8	16 tok/s 9 tok/s	~8.3s ~9.9s
#136	Qwen3.5-4B	4B	INT4 Q8 FP16	61 tok/s 36 tok/s 19 tok/s	~2.9s ~2.9s ~2.9s
#139	GPT-OSS 20B	21B (3.6B active)	INT4	20 tok/s	~3.7s
#140	GLM-4V	9B	INT4 Q8	20 tok/s 13 tok/s	~6.3s ~6.9s
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	9 tok/s 5 tok/s 3 tok/s	~19.9s ~20.0s ~20.1s
#152	Granite 4.2 8B	8B	INT4 Q8	33 tok/s 19 tok/s	~5.6s ~5.6s
#157	Gemma 4 E4B	8B	INT4 Q8	33 tok/s 19 tok/s	~5.6s ~5.6s
#163	Gemma 3 27B	27B	INT4	9 tok/s	~21.8s
#164	Phi-4	14B	INT4 Q8	19 tok/s 10 tok/s	~9.7s ~11.5s
#176	DeepSeek-R1 14B	14B	INT4 Q8	20 tok/s 10 tok/s	~9.7s ~11.5s
#177	Ministral 3 14B	14B	INT4	14 tok/s	~9.8s
#180	Gemma 3 12B	12B	INT4 Q8	22 tok/s 12 tok/s	~8.4s ~9.1s
#184	Mistral-Small-2501	24B	INT4	11 tok/s	~17.9s
#185	Ministral 3 8B	8B	INT4 Q8	23 tok/s 15 tok/s	~5.6s ~5.6s
#186	Gemma 2 27B	27B	INT4	9 tok/s	~21.8s
#187	OLMo 3 7B Think	7B	INT4 Q8	25 tok/s 17 tok/s	~4.9s ~5.0s
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	106 tok/s 66 tok/s 36 tok/s	~1.5s ~1.5s ~1.5s
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	106 tok/s 66 tok/s 36 tok/s	~1.5s ~1.5s ~1.5s
#190	Qwen2.5-14B	14B	INT4 Q8	19 tok/s 10 tok/s	~9.7s ~11.5s
#193	Gemma 4 E2B	5.1B	INT4 Q8 FP16	49 tok/s 29 tok/s 15 tok/s	~3.6s ~3.6s ~3.7s
#195	Ministral 3 3B	3B	INT4 Q8 FP16	46 tok/s 34 tok/s 21 tok/s	~2.2s ~2.2s ~2.2s
#196	DeepSeek-R1 8B	8B	INT4 Q8	34 tok/s 19 tok/s	~5.6s ~5.6s
#198	Gemma 2 9B	9B	INT4 Q8	29 tok/s 17 tok/s	~6.3s ~6.3s
#201	Qwen3-4B	4B	INT4 Q8 FP16	61 tok/s 36 tok/s 19 tok/s	~2.9s ~2.9s ~2.9s
#202	Qwen3-14B	14B	INT4 Q8	19 tok/s 10 tok/s	~9.7s ~11.5s
#203	Gemma 3 4B	4B	INT4 Q8 FP16	61 tok/s 36 tok/s 19 tok/s	~2.9s ~2.9s ~2.9s
#204	DeepSeek-R1 7B	7B	INT4 Q8 FP16	38 tok/s 22 tok/s 11 tok/s	~4.9s ~4.9s ~5.4s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	213 tok/s 144 tok/s 84 tok/s	~620 ms ~622 ms ~627 ms
#207	Llama 3.1 8B	8B	INT4 Q8	33 tok/s 19 tok/s	~5.6s ~5.6s
#207	Qwen3-8B	8B	INT4 Q8	33 tok/s 19 tok/s	~5.6s ~5.6s
#212	Llama 3 8B	8B	INT4 Q8	33 tok/s 19 tok/s	~5.6s ~5.6s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	121 tok/s 77 tok/s 42 tok/s	~1.3s ~1.3s ~1.3s
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	299 tok/s 247 tok/s 175 tok/s	~203 ms ~204 ms ~205 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	259 tok/s 180 tok/s 108 tok/s	~457 ms ~459 ms ~463 ms
#217	Phi-4-Mini	3.8B	INT4 Q8 FP16	63 tok/s 38 tok/s 20 tok/s	~2.8s ~2.8s ~2.8s
#218	Qwen2.5-7B	7B	INT4 Q8 FP16	37 tok/s 21 tok/s 10 tok/s	~4.9s ~5.0s ~5.9s
#220	Ministral-8B-2410	8B	INT4 Q8	33 tok/s 19 tok/s	~5.6s ~5.6s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	133 tok/s 85 tok/s 47 tok/s	~1.2s ~1.2s ~1.2s
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
#223	Phi-3-mini	3.8B	INT4 Q8 FP16	63 tok/s 38 tok/s 20 tok/s	~2.8s ~2.8s ~2.8s
#224	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	46 tok/s 38 tok/s 24 tok/s	~1.6s ~1.7s ~1.8s
#225	Gemma 2 2B	2B	INT4 Q8 FP16	106 tok/s 66 tok/s 36 tok/s	~1.5s ~1.5s ~1.5s
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	36 tok/s 20 tok/s 10 tok/s	~5.1s ~5.2s ~6.1s
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	179 tok/s 119 tok/s 68 tok/s	~789 ms ~792 ms ~798 ms
#228	Phi-3-medium	14B	INT4 Q8	19 tok/s 10 tok/s	~9.7s ~11.5s
#230	OLMo 3 7B Instruct	7B	INT4 Q8	25 tok/s 17 tok/s	~4.9s ~5.0s
#231	Qwen2-7B	7B	INT4 Q8 FP16	37 tok/s 21 tok/s 10 tok/s	~4.9s ~5.0s ~5.9s
#232	Phi-3-small	7B	INT4 Q8 FP16	37 tok/s 21 tok/s 10 tok/s	~4.9s ~5.0s ~5.9s
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	36 tok/s 20 tok/s 10 tok/s	~5.1s ~5.2s ~6.1s
#234	Gemma 3 1B	1B	INT4 Q8 FP16	179 tok/s 119 tok/s 68 tok/s	~789 ms ~792 ms ~798 ms
#236	Gemma 1 7B	7B	INT4 Q8	25 tok/s 17 tok/s	~4.9s ~5.0s
#237	Gemma 1 2B	2B	INT4 Q8 FP16	117 tok/s 70 tok/s 37 tok/s	~1.5s ~1.5s ~1.5s
#238	ChatGLM3-6B	6B	INT4 Q8 FP16	28 tok/s 19 tok/s 11 tok/s	~4.3s ~4.3s ~4.7s
#239	ChatGLM2-6B	6B	INT4 Q8 FP16	28 tok/s 19 tok/s 11 tok/s	~4.3s ~4.3s ~4.7s
#240	ChatGLM-6B	6B	INT4 Q8 FP16	28 tok/s 19 tok/s 11 tok/s	~4.3s ~4.3s ~4.7s
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	377 tok/s 286 tok/s 186 tok/s	~226 ms ~226 ms ~228 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	290 tok/s 206 tok/s 125 tok/s	~378 ms ~380 ms ~383 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	290 tok/s 206 tok/s 125 tok/s	~378 ms ~380 ms ~383 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	278 tok/s 229 tok/s 160 tok/s	~227 ms ~227 ms ~229 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	278 tok/s 229 tok/s 160 tok/s	~227 ms ~227 ms ~229 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	259 tok/s 180 tok/s 108 tok/s	~457 ms ~459 ms ~463 ms
-	Falcon-1B	1B	INT4 Q8 FP16	198 tok/s 127 tok/s 71 tok/s	~788 ms ~791 ms ~797 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	179 tok/s 119 tok/s 68 tok/s	~789 ms ~792 ms ~798 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	179 tok/s 119 tok/s 68 tok/s	~789 ms ~792 ms ~798 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	133 tok/s 85 tok/s 47 tok/s	~1.2s ~1.2s ~1.2s
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	133 tok/s 85 tok/s 47 tok/s	~1.2s ~1.2s ~1.2s
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	121 tok/s 77 tok/s 42 tok/s	~1.3s ~1.3s ~1.3s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	116 tok/s 73 tok/s 40 tok/s	~1.4s ~1.4s ~1.4s
-	Falcon-3B	3B	INT4 Q8 FP16	83 tok/s 49 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	81 tok/s 65 tok/s 42 tok/s	~1.0s ~1.0s ~1.0s
-	Phi-1	1.3B	INT4 Q8 FP16	81 tok/s 65 tok/s 42 tok/s	~1.0s ~1.0s ~1.0s
-	Phi-1.5	1.3B	INT4 Q8 FP16	81 tok/s 65 tok/s 42 tok/s	~1.0s ~1.0s ~1.0s
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	79 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4 Q8	79 tok/s 48 tok/s	~2.2s ~2.3s
-	Falcon3-3B	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	Liquid D1 3B	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	77 tok/s 47 tok/s 25 tok/s	~2.2s ~2.2s ~2.2s
-	LFM2.5 24B A2B	24B (2.5B active)	INT4	67 tok/s	~2.7s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	61 tok/s	~2.9s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	61 tok/s	~2.9s
-	FrogNano 4B	4B	INT4 Q8 FP16	61 tok/s 36 tok/s 19 tok/s	~2.9s ~2.9s ~2.9s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	61 tok/s 36 tok/s 19 tok/s	~2.9s ~2.9s ~2.9s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	57 tok/s 45 tok/s 29 tok/s	~1.5s ~1.5s ~1.6s
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4	55 tok/s	~3.3s
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4	55 tok/s	~3.3s
-	Phi-2	2.7B	INT4 Q8 FP16	49 tok/s 37 tok/s 23 tok/s	~2.0s ~2.0s ~2.0s
-	MaLLaM-3B	3B	INT4 Q8 FP16	46 tok/s 34 tok/s 21 tok/s	~2.2s ~2.2s ~2.2s
-	SmolLM3 3B	3B	INT4 Q8 FP16	46 tok/s 34 tok/s 21 tok/s	~2.2s ~2.2s ~2.2s
-	Falcon-7B	7B	INT4 Q8 FP16	39 tok/s 22 tok/s 11 tok/s	~4.9s ~4.9s ~5.4s
-	Falcon3-7B	7B	INT4 Q8 FP16	37 tok/s 21 tok/s 10 tok/s	~4.9s ~5.0s ~5.9s
-	Hy-MT2 7B	7B	INT4 Q8 FP16	37 tok/s 21 tok/s 10 tok/s	~4.9s ~5.0s ~5.9s
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	36 tok/s 20 tok/s 10 tok/s	~5.1s ~5.2s ~6.1s
-	Youtu Parsing Omni	8B	INT4 Q8	34 tok/s 19 tok/s	~5.6s ~5.6s
-	LensVLM 9B	9B	INT4 Q8	29 tok/s 17 tok/s	~6.3s ~6.3s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	29 tok/s 17 tok/s	~6.3s ~6.3s
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	28 tok/s 19 tok/s 11 tok/s	~4.3s ~4.3s ~4.7s
-	Yi-6B	6B	INT4 Q8 FP16	28 tok/s 19 tok/s 11 tok/s	~4.3s ~4.3s ~4.7s
-	Falcon3-10B	10B	INT4 Q8	26 tok/s 15 tok/s	~7.0s ~7.0s
-	Falcon2-11B	11B	INT4 Q8	25 tok/s 14 tok/s	~7.7s ~7.7s
-	Hunyuan Lite	7B	INT4 Q8	25 tok/s 17 tok/s	~4.9s ~5.0s
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4	25 tok/s	~2.8s
-	OLMo 3 7B Base	7B	INT4 Q8	25 tok/s 17 tok/s	~4.9s ~5.0s
-	SEA-LION-7B	7.1B	INT4 Q8	25 tok/s 17 tok/s	~5.0s ~5.0s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	25 tok/s 17 tok/s	~5.0s ~5.0s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	23 tok/s 15 tok/s	~5.6s ~5.6s
-	Typhoon-2-8B	8B	INT4 Q8	23 tok/s 15 tok/s	~5.6s ~5.6s
-	GLM-4-9B	9B	INT4 Q8	20 tok/s 13 tok/s	~6.3s ~6.9s
-	GLM-4-9B-Chat	9B	INT4 Q8	20 tok/s 13 tok/s	~6.3s ~6.9s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	20 tok/s 13 tok/s	~6.3s ~6.9s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8	20 tok/s 13 tok/s	~6.5s ~7.0s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8	20 tok/s 13 tok/s	~6.5s ~7.0s
-	Yi-9B	9B	INT4 Q8	20 tok/s 13 tok/s	~6.3s ~6.9s
-	GreenMind-14B-R1	14B	INT4	14 tok/s	~9.8s
-	AREX-2	27B	INT4	9 tok/s	~21.8s
-	PPLX Decider v1 27B	27B	INT4	9 tok/s	~21.8s
-	PPLX Decider v1.1 27B	27B	INT4	9 tok/s	~21.8s

NVIDIA A2 (16GB)

VRAM Scaling by Context Length

Best Models to Run on NVIDIA A2 (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions