What are the best models to run on AMD RX 6900 XT (16GB)?

Based on model rankings and 16 GB usable VRAM: Top Overall: Qwen 3.8 27B (INT4) ~24 tok/s. Top Coding: Qwen3.5-27B (INT4) ~24 tok/s. Optimal Daily Driver: Phi-4 Reasoning Plus (INT4) ~34 tok/s. Maximum Model Size: ERNIE-4.5-VL-28B-A3B (INT4) ~111 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 6900 XT (16GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#33	Qwen 3.8 27B	27B	INT4	24 tok/s	~7.5s
#83	Qwen3.5-27B	27B	INT4	24 tok/s	~7.5s
#100	Phi-4 Reasoning Plus	14B	INT4	34 tok/s	~3.4s
#101	MiMo V2 Flash	15B (309B active)	INT4	2 tok/s	~75.6s
#108	Gemma 4 26B A4B	25.2B (3.8B active)	INT4	105 tok/s	~1.2s
#109	Gemma 4 12B	11.95B	INT4 Q8	39 tok/s 24 tok/s	~2.9s ~3.4s
#113	Qwen3.5-9B	9B	INT4 Q8	68 tok/s 41 tok/s	~2.2s ~2.2s
#118	K2 Horizon 7B	7B	INT4 Q8 FP16	83 tok/s 52 tok/s 25 tok/s	~1.7s ~1.7s ~2.0s
#128	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	23 tok/s 13 tok/s 7 tok/s	~6.8s ~6.9s ~6.9s
#132	Qwen3.5-4B	4B	INT4 Q8 FP16	122 tok/s 81 tok/s 46 tok/s	~997 ms ~1.0s ~1.0s
#136	GPT-OSS 20B	21B (3.6B active)	INT4	49 tok/s	~1.3s
#146	Granite 4.2 8B	8B	INT4 Q8	75 tok/s 46 tok/s	~1.9s ~1.9s
#150	Gemma 4 E4B	8B	INT4 Q8	75 tok/s 46 tok/s	~1.9s ~1.9s
#158	GLM-4V	9B	INT4 Q8	49 tok/s 33 tok/s	~2.2s ~2.4s
#159	Gemma 3 27B	27B	INT4	24 tok/s	~7.5s
#160	Phi-4	14B	INT4 Q8	47 tok/s 25 tok/s	~3.3s ~4.0s
#169	OLMo 3 7B Think	7B	INT4 Q8	60 tok/s 42 tok/s	~1.7s ~1.7s
#172	DeepSeek-R1 14B	14B	INT4 Q8	48 tok/s 25 tok/s	~3.3s ~3.9s
#173	Ministral 3 14B	14B	INT4	34 tok/s	~3.4s
#177	Gemma 3 12B	12B	INT4 Q8	54 tok/s 30 tok/s	~2.9s ~3.1s
#181	Mistral-Small-2501	24B	INT4	28 tok/s	~6.2s
#182	Ministral 3 8B	8B	INT4 Q8	54 tok/s 38 tok/s	~1.9s ~1.9s
#183	Qwen3.5-2B	2B	INT4 Q8 FP16	178 tok/s 130 tok/s 81 tok/s	~532 ms ~534 ms ~538 ms
#184	MiniCPM5-2B	2B	INT4 Q8 FP16	178 tok/s 130 tok/s 81 tok/s	~532 ms ~534 ms ~538 ms
#185	Qwen2.5-14B	14B	INT4 Q8	47 tok/s 25 tok/s	~3.3s ~4.0s
#186	Gemma 2 27B	27B	INT4	24 tok/s	~7.5s
#188	Gemma 4 E2B	5.1B	INT4 Q8 FP16	104 tok/s 67 tok/s 38 tok/s	~1.3s ~1.3s ~1.3s
#189	Qwen3-4B	4B	INT4 Q8 FP16	122 tok/s 81 tok/s 46 tok/s	~997 ms ~1.0s ~1.0s
#190	DeepSeek-R1 8B	8B	INT4 Q8	76 tok/s 47 tok/s	~1.9s ~1.9s
#193	Ministral 3 3B	3B	INT4 Q8 FP16	99 tok/s 77 tok/s 50 tok/s	~767 ms ~769 ms ~776 ms
#195	Qwen3-14B	14B	INT4 Q8	47 tok/s 25 tok/s	~3.3s ~4.0s
#197	Gemma 2 9B	9B	INT4 Q8	68 tok/s 41 tok/s	~2.2s ~2.2s
#198	Qwen2.5-7B	7B	INT4 Q8 FP16	83 tok/s 52 tok/s 25 tok/s	~1.7s ~1.7s ~2.0s
#199	Qwen3-8B	8B	INT4 Q8	75 tok/s 46 tok/s	~1.9s ~1.9s
#200	Gemma 3 4B	4B	INT4 Q8 FP16	122 tok/s 81 tok/s 46 tok/s	~997 ms ~1.0s ~1.0s
#201	DeepSeek-R1 7B	7B	INT4 Q8 FP16	85 tok/s 52 tok/s 27 tok/s	~1.7s ~1.7s ~1.9s
#202	Llama 3.1 8B	8B	INT4 Q8	75 tok/s 46 tok/s	~1.9s ~1.9s
#204	Qwen3-1.7B	1.7B	INT4 Q8 FP16	193 tok/s 144 tok/s 92 tok/s	~455 ms ~457 ms ~461 ms
#206	Gemma 3 270M	0.27B	INT4 Q8 FP16	296 tok/s 275 tok/s 235 tok/s	~76 ms ~76 ms ~77 ms
#208	Llama 3 8B	8B	INT4 Q8	75 tok/s 46 tok/s	~1.9s ~1.9s
#209	Qwen3-0.6B	0.6B	INT4 Q8 FP16	280 tok/s 239 tok/s 180 tok/s	~163 ms ~164 ms ~165 ms
#210	Phi-4-Mini	3.8B	INT4 Q8 FP16	126 tok/s 84 tok/s 48 tok/s	~950 ms ~954 ms ~963 ms
#212	Ministral-8B-2410	8B	INT4 Q8	75 tok/s 46 tok/s	~1.9s ~1.9s
#213	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	204 tok/s 155 tok/s 101 tok/s	~403 ms ~405 ms ~408 ms
#214	Llama 3.2 3B	3B	INT4 Q8 FP16	145 tok/s 100 tok/s 59 tok/s	~763 ms ~766 ms ~773 ms
#215	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	258 tok/s 213 tok/s 153 tok/s	~219 ms ~220 ms ~222 ms
#216	Phi-3-mini	3.8B	INT4 Q8 FP16	126 tok/s 84 tok/s 48 tok/s	~950 ms ~954 ms ~963 ms
#217	Gemma 3n E2B IT	6B (2B active)	INT4 Q8 FP16	99 tok/s 84 tok/s 58 tok/s	~571 ms ~573 ms ~627 ms
#218	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8 FP16	80 tok/s 50 tok/s 24 tok/s	~1.8s ~1.8s ~2.1s
#219	Llama 3.2 1B	1B	INT4 Q8 FP16	238 tok/s 191 tok/s 133 tok/s	~277 ms ~278 ms ~280 ms
#220	Gemma 2 2B	2B	INT4 Q8 FP16	178 tok/s 130 tok/s 81 tok/s	~532 ms ~534 ms ~538 ms
#221	Phi-3-medium	14B	INT4 Q8	47 tok/s 25 tok/s	~3.3s ~4.0s
#222	OLMo 3 7B Instruct	7B	INT4 Q8	60 tok/s 42 tok/s	~1.7s ~1.7s
#224	Qwen2-7B	7B	INT4 Q8 FP16	83 tok/s 52 tok/s 25 tok/s	~1.7s ~1.7s ~2.0s
#225	Phi-3-small	7B	INT4 Q8 FP16	83 tok/s 52 tok/s 25 tok/s	~1.7s ~1.7s ~2.0s
#226	Gemma 3 1B	1B	INT4 Q8 FP16	238 tok/s 191 tok/s 133 tok/s	~277 ms ~278 ms ~280 ms
#227	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8 FP16	80 tok/s 50 tok/s 24 tok/s	~1.8s ~1.8s ~2.1s
#229	Gemma 1 7B	7B	INT4 Q8	60 tok/s 42 tok/s	~1.7s ~1.7s
#230	Gemma 1 2B	2B	INT4 Q8 FP16	189 tok/s 135 tok/s 83 tok/s	~531 ms ~533 ms ~538 ms
#231	ChatGLM3-6B	6B	INT4 Q8 FP16	66 tok/s 48 tok/s 27 tok/s	~1.5s ~1.5s ~1.6s
#232	ChatGLM2-6B	6B	INT4 Q8 FP16	66 tok/s 48 tok/s 27 tok/s	~1.5s ~1.5s ~1.6s
#233	ChatGLM-6B	6B	INT4 Q8 FP16	66 tok/s 48 tok/s 27 tok/s	~1.5s ~1.5s ~1.6s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	293 tok/s 254 tok/s 197 tok/s	~136 ms ~137 ms ~138 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	293 tok/s 254 tok/s 197 tok/s	~136 ms ~137 ms ~138 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	288 tok/s 266 tok/s 225 tok/s	~84 ms ~85 ms ~85 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	288 tok/s 266 tok/s 225 tok/s	~84 ms ~85 ms ~85 ms
-	Falcon-1B	1B	INT4 Q8 FP16	249 tok/s 198 tok/s 136 tok/s	~277 ms ~278 ms ~280 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	238 tok/s 191 tok/s 133 tok/s	~277 ms ~278 ms ~280 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	204 tok/s 155 tok/s 101 tok/s	~403 ms ~405 ms ~408 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	204 tok/s 155 tok/s 101 tok/s	~403 ms ~405 ms ~408 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	193 tok/s 144 tok/s 92 tok/s	~455 ms ~457 ms ~461 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	188 tok/s 139 tok/s 88 tok/s	~479 ms ~481 ms ~485 ms
-	Falcon-3B	3B	INT4 Q8 FP16	153 tok/s 103 tok/s 60 tok/s	~763 ms ~766 ms ~773 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	150 tok/s 128 tok/s 92 tok/s	~354 ms ~355 ms ~358 ms
-	Phi-1	1.3B	INT4 Q8 FP16	150 tok/s 128 tok/s 92 tok/s	~354 ms ~355 ms ~358 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	150 tok/s 128 tok/s 92 tok/s	~354 ms ~355 ms ~358 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	147 tok/s 101 tok/s 59 tok/s	~763 ms ~766 ms ~773 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	145 tok/s 100 tok/s 59 tok/s	~763 ms ~766 ms ~773 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	145 tok/s 100 tok/s 59 tok/s	~763 ms ~766 ms ~773 ms
-	Spark X2.5 4B	4B	INT4 Q8 FP16	122 tok/s 81 tok/s 46 tok/s	~997 ms ~1.0s ~1.0s
-	ERNIE-4.5-21B-A3B	21B (3B active)	INT4	121 tok/s	~1.0s
-	ERNIE-4.5-21B-A3B-Base	21B (3B active)	INT4	121 tok/s	~1.0s
-	ERNIE-4.5-VL-28B-A3B	28B (3B active)	INT4	111 tok/s	~1.2s
-	ERNIE-4.5-VL-28B-A3B-Base	28B (3B active)	INT4	111 tok/s	~1.2s
-	Phi-2	2.7B	INT4 Q8 FP16	104 tok/s 82 tok/s 54 tok/s	~697 ms ~700 ms ~706 ms
-	MaLLaM-3B	3B	INT4 Q8 FP16	99 tok/s 77 tok/s 50 tok/s	~767 ms ~769 ms ~776 ms
-	SmolLM3 3B	3B	INT4 Q8 FP16	99 tok/s 77 tok/s 50 tok/s	~767 ms ~769 ms ~776 ms
-	Falcon-7B	7B	INT4 Q8 FP16	86 tok/s 53 tok/s 27 tok/s	~1.7s ~1.7s ~1.9s
-	Falcon3-7B	7B	INT4 Q8 FP16	83 tok/s 52 tok/s 25 tok/s	~1.7s ~1.7s ~2.0s
-	Hy-MT2 7B	7B	INT4 Q8 FP16	83 tok/s 52 tok/s 25 tok/s	~1.7s ~1.7s ~2.0s
-	Mistral-7B-v0.1	7.3B	INT4 Q8 FP16	80 tok/s 50 tok/s 24 tok/s	~1.8s ~1.8s ~2.1s
-	LensVLM 9B	9B	INT4 Q8	68 tok/s 41 tok/s	~2.2s ~2.2s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	68 tok/s 41 tok/s	~2.2s ~2.2s
-	ChatGLM3-6B-32K	6B	INT4 Q8 FP16	66 tok/s 48 tok/s 27 tok/s	~1.5s ~1.5s ~1.6s
-	Yi-6B	6B	INT4 Q8 FP16	66 tok/s 48 tok/s 27 tok/s	~1.5s ~1.5s ~1.6s
-	Falcon3-10B	10B	INT4 Q8	62 tok/s 38 tok/s	~2.4s ~2.4s
-	Falcon2-11B	11B	INT4 Q8	60 tok/s 35 tok/s	~2.6s ~2.6s
-	Hunyuan Lite	7B	INT4 Q8	60 tok/s 42 tok/s	~1.7s ~1.7s
-	OLMo 3 7B Base	7B	INT4 Q8	60 tok/s 42 tok/s	~1.7s ~1.7s
-	Kimi-VL-A3B-Instruct	16B (3B active)	INT4	59 tok/s	~981 ms
-	SEA-LION-7B	7.1B	INT4 Q8	59 tok/s 42 tok/s	~1.7s ~1.7s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	59 tok/s 42 tok/s	~1.7s ~1.7s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	54 tok/s 38 tok/s	~1.9s ~1.9s
-	Typhoon-2-8B	8B	INT4 Q8	54 tok/s 38 tok/s	~1.9s ~1.9s
-	GLM-4-9B	9B	INT4 Q8	49 tok/s 33 tok/s	~2.2s ~2.4s
-	GLM-4-9B-Chat	9B	INT4 Q8	49 tok/s 33 tok/s	~2.2s ~2.4s
-	GLM-4-9B-Chat-1M	9B	INT4 Q8	49 tok/s 33 tok/s	~2.2s ~2.4s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4 Q8	49 tok/s 32 tok/s	~2.2s ~2.4s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4 Q8	49 tok/s 32 tok/s	~2.2s ~2.4s
-	Yi-9B	9B	INT4 Q8	49 tok/s 33 tok/s	~2.2s ~2.4s
-	GreenMind-14B-R1	14B	INT4	34 tok/s	~3.4s

AMD RX 6900 XT (16GB)

VRAM Scaling by Context Length

Best Models to Run on AMD RX 6900 XT (16GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions