What are the best models to run on NVIDIA T4 (16GB)?

Based on model rankings and 16 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~15 tok/s. Top Coding: Qwen3.5-27B (INT4) ~15 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~22 tok/s. Maximum Model Size: ERNIE-4.5-VL-28B-A3B (INT4) ~84 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 T4 (16GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4	15 tok/s	~6.1s
#83	Qwen3.5-27B	27B	INT4	15 tok/s	~6.1s
#100	Phi-4 Reasoning Plus	14B	INT4	22 tok/s	~2.7s
#101	MiMo V2 Flash	15B (309B active)	INT4	1 tok/s	~61.4s
#108	Gemma 4 26B A4B	25.2B (3.8B active)	INT4	78 tok/s	~1.0s
#109	Gemma 4 12B	11.95B	INT4 Q8	25 tok/s 15 tok/s	~2.3s ~2.8s
#113	Qwen3.5-9B	9B	INT4 Q8	46 tok/s 26 tok/s	~1.8s ~1.8s
#118	K2 Horizon 7B	7B	INT4 Q8 FP16	57 tok/s 33 tok/s 16 tok/s	~1.4s ~1.4s ~1.7s
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	14 tok/s 8 tok/s 4 tok/s	~5.5s ~5.6s ~5.7s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	92 tok/s 56 tok/s 30 tok/s	~807 ms ~814 ms ~830 ms
#136	GPT-OSS 20B	21B (3.6B active)	INT4	32 tok/s	~1.0s
#146	Granite 4.2 8B	8B	INT4 Q8	51 tok/s 30 tok/s	~1.6s ~1.6s
#150	Gemma 4 E4B	8B	INT4 Q8	51 tok/s 30 tok/s	~1.6s ~1.6s
#158	GLM-4V	9B	INT4 Q8	32 tok/s 20 tok/s	~1.8s ~1.9s
#159	Gemma 3 27B	27B	INT4	15 tok/s	~6.1s
#160	Phi-4	14B	INT4 Q8	30 tok/s 16 tok/s	~2.7s ~3.2s
#169	OLMo 3 7B Think	7B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
#172	DeepSeek-R1 14B	14B	INT4 Q8	31 tok/s 16 tok/s	~2.7s ~3.2s
#173	Ministral 3 14B	14B	INT4	22 tok/s	~2.7s
#177	Gemma 3 12B	12B	INT4 Q8	35 tok/s 19 tok/s	~2.3s ~2.6s
#181	Mistral-Small-2501	24B	INT4	17 tok/s	~5.0s
#182	Ministral 3 8B	8B	INT4 Q8	35 tok/s 24 tok/s	~1.6s ~1.6s
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	155 tok/s 100 tok/s 56 tok/s	~429 ms ~433 ms ~441 ms
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	155 tok/s 100 tok/s 56 tok/s	~429 ms ~433 ms ~441 ms
#185	Qwen2.5-14B	14B	INT4 Q8	30 tok/s 16 tok/s	~2.7s ~3.2s
#186	Gemma 2 27B	27B	INT4	15 tok/s	~6.1s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	75 tok/s 45 tok/s 24 tok/s	~1.0s ~1.0s ~1.0s
#189	Qwen3-4B	4B	INT4 Q8 FP16	92 tok/s 56 tok/s 30 tok/s	~807 ms ~814 ms ~830 ms
#190	DeepSeek-R1 8B	8B	INT4 Q8	52 tok/s 30 tok/s	~1.6s ~1.6s
#193	Ministral 3 3B	3B	INT4 Q8 FP16	71 tok/s 53 tok/s 33 tok/s	~623 ms ~628 ms ~640 ms
#195	Qwen3-14B	14B	INT4 Q8	30 tok/s 16 tok/s	~2.7s ~3.2s
#197	Gemma 2 9B	9B	INT4 Q8	46 tok/s 26 tok/s	~1.8s ~1.8s
#198	Qwen2.5-7B	7B	INT4 Q8 FP16	57 tok/s 33 tok/s 16 tok/s	~1.4s ~1.4s ~1.7s
#199	Qwen3-8B	8B	INT4 Q8	51 tok/s 30 tok/s	~1.6s ~1.6s
#200	Gemma 3 4B	4B	INT4 Q8 FP16	92 tok/s 56 tok/s 30 tok/s	~807 ms ~814 ms ~830 ms
#201	DeepSeek-R1 7B	7B	INT4 Q8 FP16	59 tok/s 34 tok/s 17 tok/s	~1.4s ~1.4s ~1.5s
#202	Llama 3.1 8B	8B	INT4 Q8	51 tok/s 30 tok/s	~1.6s ~1.6s
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	174 tok/s 115 tok/s 65 tok/s	~367 ms ~370 ms ~377 ms
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	377 tok/s 324 tok/s 241 tok/s	~59 ms ~60 ms ~61 ms
#208	Llama 3 8B	8B	INT4 Q8	51 tok/s 30 tok/s	~1.6s ~1.6s
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	336 tok/s 248 tok/s 157 tok/s	~130 ms ~131 ms ~133 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	96 tok/s 58 tok/s 31 tok/s	~769 ms ~776 ms ~791 ms
#212	Ministral-8B-2410	8B	INT4 Q8	51 tok/s 30 tok/s	~1.6s ~1.6s
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	191 tok/s 127 tok/s 73 tok/s	~325 ms ~328 ms ~333 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	116 tok/s 72 tok/s 39 tok/s	~618 ms ~623 ms ~635 ms
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	286 tok/s 204 tok/s 125 tok/s	~175 ms ~177 ms ~180 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	96 tok/s 58 tok/s 31 tok/s	~769 ms ~776 ms ~791 ms
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	71 tok/s 59 tok/s 38 tok/s	~466 ms ~468 ms ~517 ms
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	55 tok/s 32 tok/s 15 tok/s	~1.4s ~1.4s ~1.7s
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	247 tok/s 172 tok/s 103 tok/s	~222 ms ~224 ms ~228 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	155 tok/s 100 tok/s 56 tok/s	~429 ms ~433 ms ~441 ms
#221	Phi-3-medium	14B	INT4 Q8	30 tok/s 16 tok/s	~2.7s ~3.2s
#222	OLMo 3 7B Instruct	7B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
#224	Qwen2-7B	7B	INT4 Q8 FP16	57 tok/s 33 tok/s 16 tok/s	~1.4s ~1.4s ~1.7s
#225	Phi-3-small	7B	INT4 Q8 FP16	57 tok/s 33 tok/s 16 tok/s	~1.4s ~1.4s ~1.7s
#226	Gemma 3 1B	1B	INT4 Q8 FP16	247 tok/s 172 tok/s 103 tok/s	~222 ms ~224 ms ~228 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	55 tok/s 32 tok/s 15 tok/s	~1.4s ~1.4s ~1.7s
#229	Gemma 1 7B	7B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
#230	Gemma 1 2B	2B	INT4 Q8 FP16	169 tok/s 106 tok/s 58 tok/s	~429 ms ~432 ms ~440 ms
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~1.2s ~1.2s ~1.3s
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~1.2s ~1.2s ~1.3s
#233	ChatGLM-6B	6B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~1.2s ~1.2s ~1.3s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	368 tok/s 278 tok/s 180 tok/s	~108 ms ~109 ms ~111 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	368 tok/s 278 tok/s 180 tok/s	~108 ms ~109 ms ~111 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	355 tok/s 304 tok/s 224 tok/s	~66 ms ~67 ms ~68 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	355 tok/s 304 tok/s 224 tok/s	~66 ms ~67 ms ~68 ms
-	Falcon-1B	1B	INT4 Q8 FP16	268 tok/s 182 tok/s 106 tok/s	~222 ms ~224 ms ~228 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	247 tok/s 172 tok/s 103 tok/s	~222 ms ~224 ms ~228 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	191 tok/s 127 tok/s 73 tok/s	~325 ms ~328 ms ~333 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	191 tok/s 127 tok/s 73 tok/s	~325 ms ~328 ms ~333 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	174 tok/s 115 tok/s 65 tok/s	~367 ms ~370 ms ~377 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	168 tok/s 110 tok/s 62 tok/s	~387 ms ~390 ms ~397 ms
-	Falcon-3B	3B	INT4 Q8 FP16	124 tok/s 75 tok/s 40 tok/s	~617 ms ~622 ms ~634 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	121 tok/s 98 tok/s 65 tok/s	~287 ms ~289 ms ~294 ms
-	Phi-1	1.3B	INT4 Q8 FP16	121 tok/s 98 tok/s 65 tok/s	~287 ms ~289 ms ~294 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	121 tok/s 98 tok/s 65 tok/s	~287 ms ~289 ms ~294 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	118 tok/s 72 tok/s 39 tok/s	~617 ms ~623 ms ~634 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	116 tok/s 72 tok/s 39 tok/s	~618 ms ~623 ms ~635 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	116 tok/s 72 tok/s 39 tok/s	~618 ms ~623 ms ~635 ms
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	93 tok/s	~820 ms
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	93 tok/s	~820 ms
-	Spark X2.5 4B	4B	INT4 Q8 FP16	92 tok/s 56 tok/s 30 tok/s	~807 ms ~814 ms ~830 ms
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4	84 tok/s	~934 ms
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4	84 tok/s	~934 ms
-	Phi-2	2.7B	INT4 Q8 FP16	75 tok/s 57 tok/s 35 tok/s	~566 ms ~571 ms ~581 ms
-	MaLLaM-3B	3B	INT4 Q8 FP16	71 tok/s 53 tok/s 33 tok/s	~623 ms ~628 ms ~640 ms
-	SmolLM3 3B	3B	INT4 Q8 FP16	71 tok/s 53 tok/s 33 tok/s	~623 ms ~628 ms ~640 ms
-	Falcon-7B	7B	INT4 Q8 FP16	60 tok/s 34 tok/s 17 tok/s	~1.4s ~1.4s ~1.5s
-	Falcon3-7B	7B	INT4 Q8 FP16	57 tok/s 33 tok/s 16 tok/s	~1.4s ~1.4s ~1.7s
-	Hy-MT2 7B	7B	INT4 Q8 FP16	57 tok/s 33 tok/s 16 tok/s	~1.4s ~1.4s ~1.7s
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	55 tok/s 32 tok/s 15 tok/s	~1.4s ~1.4s ~1.7s
-	LensVLM 9B	9B	INT4 Q8	46 tok/s 26 tok/s	~1.8s ~1.8s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	46 tok/s 26 tok/s	~1.8s ~1.8s
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~1.2s ~1.2s ~1.3s
-	Yi-6B	6B	INT4 Q8 FP16	44 tok/s 31 tok/s 17 tok/s	~1.2s ~1.2s ~1.3s
-	Falcon3-10B	10B	INT4 Q8	41 tok/s 24 tok/s	~1.9s ~2.0s
-	Falcon2-11B	11B	INT4 Q8	40 tok/s 22 tok/s	~2.1s ~2.2s
-	Hunyuan Lite	7B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4	39 tok/s	~802 ms
-	OLMo 3 7B Base	7B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
-	SEA-LION-7B	7.1B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	39 tok/s 27 tok/s	~1.4s ~1.4s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	35 tok/s 24 tok/s	~1.6s ~1.6s
-	Typhoon-2-8B	8B	INT4 Q8	35 tok/s 24 tok/s	~1.6s ~1.6s
-	GLM-4-9B	9B	INT4 Q8	32 tok/s 20 tok/s	~1.8s ~1.9s
-	GLM-4-9B-Chat	9B	INT4 Q8	32 tok/s 20 tok/s	~1.8s ~1.9s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	32 tok/s 20 tok/s	~1.8s ~1.9s
-	Yi-9B	9B	INT4 Q8	32 tok/s 20 tok/s	~1.8s ~1.9s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8	31 tok/s 20 tok/s	~1.8s ~2.0s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8	31 tok/s 20 tok/s	~1.8s ~2.0s
-	GreenMind-14B-R1	14B	INT4	22 tok/s	~2.7s

NVIDIA T4 (16GB)

VRAM Scaling by Context Length

Best Models to Run on NVIDIA T4 (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions