What are the best models to run on AMD RX 6650 XT (8GB)?

Based on model rankings and 8 GB usable VRAM: Top Overall: Qwen3.5-9B (INT4) ~40 tok/s. Top Coding: K2 Horizon 7B (INT4) ~50 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~45 tok/s. Maximum Model Size: Mellum 2.1 12B A2.5B Thinking (INT4) ~88 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 AMD RX 6650 XT (8GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#121	Qwen3.5-9B	9B	INT4	40 tok/s	~4.5s
#127	K2 Horizon 7B	7B	INT4	50 tok/s	~3.5s
#136	Qwen3.5-4B	4B	INT4 Q8	78 tok/s 49 tok/s	~2.1s ~2.1s
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	13 tok/s 7 tok/s	~14.2s ~15.5s
#152	Granite 4.2 8B	8B	INT4	45 tok/s	~4.0s
#157	Gemma 4 E4B	8B	INT4	45 tok/s	~4.0s
#180	Gemma 3 12B	12B	INT4	28 tok/s	~7.1s
#185	Ministral 3 8B	8B	INT4	29 tok/s	~4.7s
#187	OLMo 3 7B Think	7B	INT4	33 tok/s	~3.9s
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	124 tok/s 85 tok/s 49 tok/s	~1.1s ~1.1s ~1.1s
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	124 tok/s 85 tok/s 49 tok/s	~1.1s ~1.1s ~1.1s
#193	Gemma 4 E2B	5.1B	INT4 Q8	65 tok/s 38 tok/s	~2.6s ~2.8s
#195	Ministral 3 3B	3B	INT4 Q8	61 tok/s 46 tok/s	~1.6s ~1.6s
#196	DeepSeek-R1 8B	8B	INT4	46 tok/s	~4.0s
#198	Gemma 2 9B	9B	INT4	40 tok/s	~4.5s
#201	Qwen3-4B	4B	INT4 Q8	78 tok/s 49 tok/s	~2.1s ~2.1s
#203	Gemma 3 4B	4B	INT4 Q8	78 tok/s 49 tok/s	~2.1s ~2.1s
#204	DeepSeek-R1 7B	7B	INT4 Q8	52 tok/s 27 tok/s	~3.5s ~4.2s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	206 tok/s 157 tok/s 103 tok/s	~449 ms ~451 ms ~454 ms
#207	Llama 3.1 8B	8B	INT4	45 tok/s	~4.0s
#207	Qwen3-8B	8B	INT4	45 tok/s	~4.0s
#212	Llama 3 8B	8B	INT4	45 tok/s	~4.0s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	138 tok/s 96 tok/s 57 tok/s	~942 ms ~945 ms ~952 ms
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	254 tok/s 227 tok/s 180 tok/s	~151 ms ~151 ms ~152 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	233 tok/s 185 tok/s 126 tok/s	~333 ms ~334 ms ~336 ms
#217	Phi-4-Mini	3.8B	INT4 Q8	81 tok/s 51 tok/s	~2.0s ~2.0s
#218	Qwen2.5-7B	7B	INT4	50 tok/s	~3.5s
#220	Ministral-8B-2410	8B	INT4	45 tok/s	~4.0s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	149 tok/s 105 tok/s 63 tok/s	~834 ms ~837 ms ~843 ms
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
#223	Phi-3-mini	3.8B	INT4 Q8	81 tok/s 51 tok/s	~2.0s ~2.0s
#224	Gemma 3n E2B IT	6B (2B active)	INT4	58 tok/s	~1.3s
#225	Gemma 2 2B	2B	INT4 Q8 FP16	124 tok/s 85 tok/s 49 tok/s	~1.1s ~1.1s ~1.1s
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4	49 tok/s	~3.7s
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	184 tok/s 137 tok/s 87 tok/s	~570 ms ~572 ms ~576 ms
#230	OLMo 3 7B Instruct	7B	INT4	33 tok/s	~3.9s
#231	Qwen2-7B	7B	INT4	50 tok/s	~3.5s
#232	Phi-3-small	7B	INT4	50 tok/s	~3.5s
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4	49 tok/s	~3.7s
#234	Gemma 3 1B	1B	INT4 Q8 FP16	184 tok/s 137 tok/s 87 tok/s	~570 ms ~572 ms ~576 ms
#236	Gemma 1 7B	7B	INT4	33 tok/s	~3.9s
#237	Gemma 1 2B	2B	INT4 Q8 FP16	134 tok/s 89 tok/s 51 tok/s	~1.1s ~1.1s ~1.1s
#238	ChatGLM3-6B	6B	INT4	37 tok/s	~3.3s
#239	ChatGLM2-6B	6B	INT4	37 tok/s	~3.3s
#240	ChatGLM-6B	6B	INT4	37 tok/s	~3.3s
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	288 tok/s 247 tok/s 189 tok/s	~167 ms ~167 ms ~169 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	249 tok/s 202 tok/s 142 tok/s	~276 ms ~277 ms ~279 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	249 tok/s 202 tok/s 142 tok/s	~276 ms ~277 ms ~279 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	243 tok/s 216 tok/s 170 tok/s	~167 ms ~168 ms ~169 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	243 tok/s 216 tok/s 170 tok/s	~167 ms ~168 ms ~169 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	233 tok/s 185 tok/s 126 tok/s	~333 ms ~334 ms ~336 ms
-	Falcon-1B	1B	INT4 Q8 FP16	196 tok/s 143 tok/s 89 tok/s	~570 ms ~572 ms ~576 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	184 tok/s 137 tok/s 87 tok/s	~570 ms ~572 ms ~576 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	184 tok/s 137 tok/s 87 tok/s	~570 ms ~572 ms ~576 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	149 tok/s 105 tok/s 63 tok/s	~834 ms ~837 ms ~843 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	149 tok/s 105 tok/s 63 tok/s	~834 ms ~837 ms ~843 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	138 tok/s 96 tok/s 57 tok/s	~942 ms ~945 ms ~952 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	133 tok/s 92 tok/s 54 tok/s	~992 ms ~995 ms ~1.0s
-	Falcon-3B	3B	INT4 Q8 FP16	103 tok/s 65 tok/s 34 tok/s	~1.6s ~1.6s ~1.7s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	100 tok/s 83 tok/s 57 tok/s	~730 ms ~732 ms ~737 ms
-	Phi-1	1.3B	INT4 Q8 FP16	100 tok/s 83 tok/s 57 tok/s	~730 ms ~732 ms ~737 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	100 tok/s 83 tok/s 57 tok/s	~730 ms ~732 ms ~737 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	98 tok/s 63 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	Falcon3-3B	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	Liquid D1 3B	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	96 tok/s 62 tok/s 33 tok/s	~1.6s ~1.6s ~1.7s
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	88 tok/s	~1.8s
-	FrogNano 4B	4B	INT4 Q8	78 tok/s 49 tok/s	~2.1s ~2.1s
-	Spark X2.5 4B	4B	INT4 Q8	78 tok/s 49 tok/s	~2.1s ~2.1s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	75 tok/s 60 tok/s 38 tok/s	~1.1s ~1.1s ~1.2s
-	Phi-2	2.7B	INT4 Q8 FP16	65 tok/s 50 tok/s 29 tok/s	~1.4s ~1.4s ~1.7s
-	MaLLaM-3B	3B	INT4 Q8	61 tok/s 46 tok/s	~1.6s ~1.6s
-	SmolLM3 3B	3B	INT4 Q8	61 tok/s 46 tok/s	~1.6s ~1.6s
-	Falcon-7B	7B	INT4 Q8	53 tok/s 28 tok/s	~3.5s ~4.2s
-	Falcon3-7B	7B	INT4	50 tok/s	~3.5s
-	Hy-MT2 7B	7B	INT4	50 tok/s	~3.5s
-	Mistral-7B-v0.1	7.3B	INT4	49 tok/s	~3.7s
-	Youtu Parsing Omni	8B	INT4	46 tok/s	~4.0s
-	LensVLM 9B	9B	INT4	40 tok/s	~4.5s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	40 tok/s	~4.5s
-	ChatGLM3-6B-32K	6B	INT4	37 tok/s	~3.3s
-	Yi-6B	6B	INT4	37 tok/s	~3.3s
-	Falcon3-10B	10B	INT4	35 tok/s	~5.4s
-	Falcon2-11B	11B	INT4	34 tok/s	~6.0s
-	Hunyuan Lite	7B	INT4	33 tok/s	~3.9s
-	OLMo 3 7B Base	7B	INT4	33 tok/s	~3.9s
-	SEA-LION-7B	7.1B	INT4	33 tok/s	~3.9s
-	SEA-LION-7B-Instruct	7.1B	INT4	33 tok/s	~3.9s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	29 tok/s	~4.7s
-	Typhoon-2-8B	8B	INT4	29 tok/s	~4.7s

AMD RX 6650 XT (8GB)

VRAM Scaling by Context Length

Best Models to Run on AMD RX 6650 XT (8GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions