What are the best models to run on AMD RX 9070 XT (16GB)?

Based on model rankings and 16 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~29 tok/s. Top Coding: Qwen3.5-27B (INT4) ~29 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~42 tok/s. Maximum Model Size: ERNIE-4.5-VL-28B-A3B (INT4) ~128 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 AMD RX 9070 XT (16GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4	29 tok/s	~1.8s
#83	Qwen3.5-27B	27B	INT4	29 tok/s	~1.8s
#100	Phi-4 Reasoning Plus	14B	INT4	42 tok/s	~821 ms
#101	MiMo V2 Flash	15B (309B active)	INT4	3 tok/s	~18.3s
#108	Gemma 4 26B A4B	25.2B (3.8B active)	INT4	123 tok/s	~305 ms
#109	Gemma 4 12B	11.95B	INT4 Q8	48 tok/s 30 tok/s	~704 ms ~836 ms
#113	Qwen3.5-9B	9B	INT4 Q8	81 tok/s 50 tok/s	~530 ms ~538 ms
#118	K2 Horizon 7B	7B	INT4 Q8 FP16	98 tok/s 62 tok/s 31 tok/s	~417 ms ~422 ms ~509 ms
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	28 tok/s 16 tok/s 9 tok/s	~1.7s ~1.7s ~1.7s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	141 tok/s 96 tok/s 56 tok/s	~247 ms ~250 ms ~258 ms
#136	GPT-OSS 20B	21B (3.6B active)	INT4	59 tok/s	~317 ms
#146	Granite 4.2 8B	8B	INT4 Q8	89 tok/s 56 tok/s	~473 ms ~480 ms
#150	Gemma 4 E4B	8B	INT4 Q8	89 tok/s 56 tok/s	~473 ms ~480 ms
#158	GLM-4V	9B	INT4 Q8	60 tok/s 40 tok/s	~535 ms ~588 ms
#159	Gemma 3 27B	27B	INT4	29 tok/s	~1.8s
#160	Phi-4	14B	INT4 Q8	57 tok/s 31 tok/s	~815 ms ~969 ms
#169	OLMo 3 7B Think	7B	INT4 Q8	72 tok/s 51 tok/s	~420 ms ~426 ms
#172	DeepSeek-R1 14B	14B	INT4 Q8	59 tok/s 31 tok/s	~815 ms ~968 ms
#173	Ministral 3 14B	14B	INT4	42 tok/s	~821 ms
#177	Gemma 3 12B	12B	INT4 Q8	65 tok/s 37 tok/s	~702 ms ~772 ms
#181	Mistral-Small-2501	24B	INT4	34 tok/s	~1.5s
#182	Ministral 3 8B	8B	INT4 Q8	65 tok/s 46 tok/s	~477 ms ~483 ms
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	199 tok/s 150 tok/s 97 tok/s	~134 ms ~136 ms ~140 ms
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	199 tok/s 150 tok/s 97 tok/s	~134 ms ~136 ms ~140 ms
#185	Qwen2.5-14B	14B	INT4 Q8	57 tok/s 31 tok/s	~815 ms ~969 ms
#186	Gemma 2 27B	27B	INT4	29 tok/s	~1.8s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	121 tok/s 80 tok/s 46 tok/s	~309 ms ~313 ms ~323 ms
#189	Qwen3-4B	4B	INT4 Q8 FP16	141 tok/s 96 tok/s 56 tok/s	~247 ms ~250 ms ~258 ms
#190	DeepSeek-R1 8B	8B	INT4 Q8	91 tok/s 57 tok/s	~473 ms ~479 ms
#193	Ministral 3 3B	3B	INT4 Q8 FP16	116 tok/s 92 tok/s 61 tok/s	~193 ms ~195 ms ~201 ms
#195	Qwen3-14B	14B	INT4 Q8	57 tok/s 31 tok/s	~815 ms ~969 ms
#197	Gemma 2 9B	9B	INT4 Q8	81 tok/s 50 tok/s	~530 ms ~538 ms
#198	Qwen2.5-7B	7B	INT4 Q8 FP16	98 tok/s 62 tok/s 31 tok/s	~417 ms ~422 ms ~509 ms
#199	Qwen3-8B	8B	INT4 Q8	89 tok/s 56 tok/s	~473 ms ~480 ms
#200	Gemma 3 4B	4B	INT4 Q8 FP16	141 tok/s 96 tok/s 56 tok/s	~247 ms ~250 ms ~258 ms
#201	DeepSeek-R1 7B	7B	INT4 Q8 FP16	100 tok/s 63 tok/s 33 tok/s	~416 ms ~422 ms ~471 ms
#202	Llama 3.1 8B	8B	INT4 Q8	89 tok/s 56 tok/s	~473 ms ~480 ms
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	213 tok/s 164 tok/s 108 tok/s	~116 ms ~117 ms ~120 ms
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	308 tok/s 290 tok/s 254 tok/s	~24 ms ~25 ms ~25 ms
#208	Llama 3 8B	8B	INT4 Q8	89 tok/s 56 tok/s	~473 ms ~480 ms
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	295 tok/s 257 tok/s 200 tok/s	~45 ms ~46 ms ~47 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	145 tok/s 99 tok/s 59 tok/s	~236 ms ~239 ms ~246 ms
#212	Ministral-8B-2410	8B	INT4 Q8	89 tok/s 56 tok/s	~473 ms ~480 ms
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	224 tok/s 175 tok/s 118 tok/s	~103 ms ~105 ms ~107 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	165 tok/s 117 tok/s 71 tok/s	~190 ms ~193 ms ~198 ms
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	275 tok/s 232 tok/s 173 tok/s	~59 ms ~59 ms ~61 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	145 tok/s 99 tok/s 59 tok/s	~236 ms ~239 ms ~246 ms
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	116 tok/s 100 tok/s 69 tok/s	~146 ms ~148 ms ~164 ms
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	95 tok/s 60 tok/s 30 tok/s	~433 ms ~439 ms ~529 ms
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	256 tok/s 212 tok/s 152 tok/s	~73 ms ~74 ms ~76 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	199 tok/s 150 tok/s 97 tok/s	~134 ms ~136 ms ~140 ms
#221	Phi-3-medium	14B	INT4 Q8	57 tok/s 31 tok/s	~815 ms ~969 ms
#222	OLMo 3 7B Instruct	7B	INT4 Q8	72 tok/s 51 tok/s	~420 ms ~426 ms
#224	Qwen2-7B	7B	INT4 Q8 FP16	98 tok/s 62 tok/s 31 tok/s	~417 ms ~422 ms ~509 ms
#225	Phi-3-small	7B	INT4 Q8 FP16	98 tok/s 62 tok/s 31 tok/s	~417 ms ~422 ms ~509 ms
#226	Gemma 3 1B	1B	INT4 Q8 FP16	256 tok/s 212 tok/s 152 tok/s	~73 ms ~74 ms ~76 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	95 tok/s 60 tok/s 30 tok/s	~433 ms ~439 ms ~529 ms
#229	Gemma 1 7B	7B	INT4 Q8	72 tok/s 51 tok/s	~420 ms ~426 ms
#230	Gemma 1 2B	2B	INT4 Q8 FP16	209 tok/s 155 tok/s 99 tok/s	~134 ms ~136 ms ~139 ms
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	79 tok/s 58 tok/s 33 tok/s	~363 ms ~368 ms ~411 ms
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	79 tok/s 58 tok/s 33 tok/s	~363 ms ~368 ms ~411 ms
#233	ChatGLM-6B	6B	INT4 Q8 FP16	79 tok/s 58 tok/s 33 tok/s	~363 ms ~368 ms ~411 ms
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	306 tok/s 271 tok/s 217 tok/s	~39 ms ~39 ms ~40 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	306 tok/s 271 tok/s 217 tok/s	~39 ms ~39 ms ~40 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	301 tok/s 282 tok/s 244 tok/s	~26 ms ~27 ms ~27 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	301 tok/s 282 tok/s 244 tok/s	~26 ms ~27 ms ~27 ms
-	Falcon-1B	1B	INT4 Q8 FP16	267 tok/s 218 tok/s 156 tok/s	~73 ms ~74 ms ~75 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	256 tok/s 212 tok/s 152 tok/s	~73 ms ~74 ms ~76 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	224 tok/s 175 tok/s 118 tok/s	~103 ms ~105 ms ~107 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	224 tok/s 175 tok/s 118 tok/s	~103 ms ~105 ms ~107 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	213 tok/s 164 tok/s 108 tok/s	~116 ms ~117 ms ~120 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	208 tok/s 159 tok/s 104 tok/s	~122 ms ~123 ms ~126 ms
-	Falcon-3B	3B	INT4 Q8 FP16	173 tok/s 121 tok/s 72 tok/s	~190 ms ~193 ms ~198 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	170 tok/s 147 tok/s 109 tok/s	~93 ms ~93 ms ~96 ms
-	Phi-1	1.3B	INT4 Q8 FP16	170 tok/s 147 tok/s 109 tok/s	~93 ms ~93 ms ~96 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	170 tok/s 147 tok/s 109 tok/s	~93 ms ~93 ms ~96 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	167 tok/s 118 tok/s 71 tok/s	~190 ms ~193 ms ~198 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	165 tok/s 117 tok/s 71 tok/s	~190 ms ~193 ms ~198 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	165 tok/s 117 tok/s 71 tok/s	~190 ms ~193 ms ~198 ms
-	Spark X2.5 4B	4B	INT4 Q8 FP16	141 tok/s 96 tok/s 56 tok/s	~247 ms ~250 ms ~258 ms
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	139 tok/s	~250 ms
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	139 tok/s	~250 ms
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4	128 tok/s	~284 ms
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4	128 tok/s	~284 ms
-	Phi-2	2.7B	INT4 Q8 FP16	121 tok/s 97 tok/s 66 tok/s	~176 ms ~178 ms ~183 ms
-	MaLLaM-3B	3B	INT4 Q8 FP16	116 tok/s 92 tok/s 61 tok/s	~193 ms ~195 ms ~201 ms
-	SmolLM3 3B	3B	INT4 Q8 FP16	116 tok/s 92 tok/s 61 tok/s	~193 ms ~195 ms ~201 ms
-	Falcon-7B	7B	INT4 Q8 FP16	102 tok/s 64 tok/s 33 tok/s	~416 ms ~422 ms ~471 ms
-	Falcon3-7B	7B	INT4 Q8 FP16	98 tok/s 62 tok/s 31 tok/s	~417 ms ~422 ms ~509 ms
-	Hy-MT2 7B	7B	INT4 Q8 FP16	98 tok/s 62 tok/s 31 tok/s	~417 ms ~422 ms ~509 ms
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	95 tok/s 60 tok/s 30 tok/s	~433 ms ~439 ms ~529 ms
-	LensVLM 9B	9B	INT4 Q8	81 tok/s 50 tok/s	~530 ms ~538 ms
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	81 tok/s 50 tok/s	~530 ms ~538 ms
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	79 tok/s 58 tok/s 33 tok/s	~363 ms ~368 ms ~411 ms
-	Yi-6B	6B	INT4 Q8 FP16	79 tok/s 58 tok/s 33 tok/s	~363 ms ~368 ms ~411 ms
-	Falcon3-10B	10B	INT4 Q8	75 tok/s 46 tok/s	~587 ms ~595 ms
-	Falcon2-11B	11B	INT4 Q8	72 tok/s 43 tok/s	~643 ms ~652 ms
-	Hunyuan Lite	7B	INT4 Q8	72 tok/s 51 tok/s	~420 ms ~426 ms
-	OLMo 3 7B Base	7B	INT4 Q8	72 tok/s 51 tok/s	~420 ms ~426 ms
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4	71 tok/s	~247 ms
-	SEA-LION-7B	7.1B	INT4 Q8	71 tok/s 51 tok/s	~426 ms ~431 ms
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	71 tok/s 51 tok/s	~426 ms ~431 ms
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	65 tok/s 46 tok/s	~477 ms ~483 ms
-	Typhoon-2-8B	8B	INT4 Q8	65 tok/s 46 tok/s	~477 ms ~483 ms
-	GLM-4-9B	9B	INT4 Q8	60 tok/s 40 tok/s	~535 ms ~588 ms
-	GLM-4-9B-Chat	9B	INT4 Q8	60 tok/s 40 tok/s	~535 ms ~588 ms
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	60 tok/s 40 tok/s	~535 ms ~588 ms
-	Yi-9B	9B	INT4 Q8	60 tok/s 40 tok/s	~535 ms ~588 ms
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8	59 tok/s 39 tok/s	~546 ms ~600 ms
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8	59 tok/s 39 tok/s	~546 ms ~600 ms
-	GreenMind-14B-R1	14B	INT4	42 tok/s	~821 ms

AMD RX 9070 XT (16GB)

VRAM Scaling by Context Length

Best Models to Run on AMD RX 9070 XT (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions