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) ~39 tok/s. Top Coding: K2 Horizon 7B (INT4) ~48 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~43 tok/s. Maximum Model Size: HunyuanVideo 1.5 720P T2V (INT4) ~26 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
#125	Qwen3.5-9B	9B	INT4	39 tok/s	~5.2s
#128	K2 Horizon 7B	7B	INT4	48 tok/s	~4.0s
#140	Qwen3.5-4B	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
#150	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	13 tok/s 7 tok/s	~16.3s ~17.7s
#158	Granite 4.2 8B	8B	INT4	43 tok/s	~4.6s
#159	Gemma 4 E4B	8B	INT4	43 tok/s	~4.6s
#178	OLMo 3 7B Think	7B	INT4	32 tok/s	~4.4s
#186	Gemma 3 12B	12B	INT4	28 tok/s	~8.1s
#189	Ministral 3 8B	8B	INT4	28 tok/s	~5.4s
#191	MiniCPM5-2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
#192	Qwen3.5-2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
#200	Gemma 4 E2B	5.1B	INT4 Q8	62 tok/s 37 tok/s	~3.0s ~3.2s
#202	DeepSeek-R1 8B	8B	INT4	44 tok/s	~4.6s
#204	Qwen3-4B	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
#205	Ministral 3 3B	3B	INT4 Q8	59 tok/s 45 tok/s	~1.8s ~1.8s
#207	Gemma 2 9B	9B	INT4	39 tok/s	~5.2s
#208	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	177 tok/s 140 tok/s 95 tok/s	~514 ms ~515 ms ~518 ms
#210	Gemma 3 4B	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
#211	Qwen3-8B	8B	INT4	43 tok/s	~4.6s
#213	DeepSeek-R1 7B	7B	INT4 Q8	50 tok/s 27 tok/s	~4.0s ~4.8s
#214	Qwen3-1.7B	1.7B	INT4 Q8 FP16	124 tok/s 89 tok/s 54 tok/s	~1.1s ~1.1s ~1.1s
#215	Llama 3.1 8B	8B	INT4	43 tok/s	~4.6s
#216	Gemma 3 270M	0.27B	INT4 Q8 FP16	211 tok/s 192 tok/s 158 tok/s	~173 ms ~173 ms ~174 ms
#218	Qwen3-0.6B	0.6B	INT4 Q8 FP16	197 tok/s 161 tok/s 114 tok/s	~381 ms ~382 ms ~384 ms
#219	Llama 3 8B	8B	INT4	43 tok/s	~4.6s
#221	Phi-4-Mini	3.8B	INT4 Q8	76 tok/s 49 tok/s	~2.3s ~2.3s
#223	Qwen2.5-7B	7B	INT4	48 tok/s	~4.0s
#224	Ministral-8B-2410	8B	INT4	43 tok/s	~4.6s
#225	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	133 tok/s 97 tok/s 60 tok/s	~953 ms ~956 ms ~962 ms
#226	Llama 3.2 3B	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
#228	Phi-3-mini	3.8B	INT4 Q8	76 tok/s 49 tok/s	~2.3s ~2.3s
#229	Gemma 3n E2B IT	6B (2B active)	INT4	56 tok/s	~1.5s
#230	Mistral-7B-Instruct-v0.1	7.3B	INT4	47 tok/s	~4.2s
#231	Llama 3.2 1B	1B	INT4 Q8 FP16	160 tok/s 123 tok/s 81 tok/s	~652 ms ~654 ms ~658 ms
#232	Gemma 2 2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
#234	OLMo 3 7B Instruct	7B	INT4	32 tok/s	~4.4s
#236	Qwen2-7B	7B	INT4	48 tok/s	~4.0s
#237	Phi-3-small	7B	INT4	48 tok/s	~4.0s
#238	Gemma 3 1B	1B	INT4 Q8 FP16	160 tok/s 123 tok/s 81 tok/s	~652 ms ~654 ms ~658 ms
#239	Mistral-7B-Instruct-v0.2	7.3B	INT4	47 tok/s	~4.2s
#241	Gemma 1 7B	7B	INT4	32 tok/s	~4.4s
#242	Gemma 1 2B	2B	INT4 Q8 FP16	121 tok/s 83 tok/s 49 tok/s	~1.3s ~1.3s ~1.3s
#243	ChatGLM3-6B	6B	INT4	36 tok/s	~3.8s
#244	ChatGLM2-6B	6B	INT4	36 tok/s	~3.8s
#245	ChatGLM-6B	6B	INT4	36 tok/s	~3.8s
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	234 tok/s 206 tok/s 164 tok/s	~191 ms ~192 ms ~193 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	208 tok/s 174 tok/s 127 tok/s	~316 ms ~317 ms ~319 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	208 tok/s 174 tok/s 127 tok/s	~316 ms ~317 ms ~319 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	204 tok/s 184 tok/s 149 tok/s	~192 ms ~192 ms ~194 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	204 tok/s 184 tok/s 149 tok/s	~192 ms ~192 ms ~194 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	197 tok/s 161 tok/s 114 tok/s	~381 ms ~382 ms ~384 ms
-	PixArt-α XL 2 1024-MS	0.61B	INT4 Q8 FP16	196 tok/s 160 tok/s 113 tok/s	~386 ms ~388 ms ~390 ms
-	PixArt-Σ XL 2 1024-MS	0.61B	INT4 Q8 FP16	196 tok/s 160 tok/s 113 tok/s	~386 ms ~388 ms ~390 ms
-	Nemotron 3.5 Parse	0.63B	INT4 Q8 FP16	193 tok/s 157 tok/s 111 tok/s	~401 ms ~403 ms ~405 ms
-	SD 2.1 Turbo	0.86B	INT4 Q8 FP16	172 tok/s 134 tok/s 90 tok/s	~554 ms ~555 ms ~559 ms
-	Stable Diffusion 1.4	0.86B	INT4 Q8 FP16	172 tok/s 134 tok/s 90 tok/s	~554 ms ~555 ms ~559 ms
-	Stable Diffusion 2.1	0.86B	INT4 Q8 FP16	172 tok/s 134 tok/s 90 tok/s	~554 ms ~555 ms ~559 ms
-	Falcon-1B	1B	INT4 Q8 FP16	170 tok/s 128 tok/s 83 tok/s	~651 ms ~653 ms ~658 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	160 tok/s 123 tok/s 81 tok/s	~652 ms ~654 ms ~658 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	160 tok/s 123 tok/s 81 tok/s	~652 ms ~654 ms ~658 ms
-	Kandinsky 2.2 Decoder	1.22B	INT4 Q8 FP16	147 tok/s 110 tok/s 70 tok/s	~785 ms ~787 ms ~792 ms
-	Wan 2.1 T2V 1.3B	1.3B	INT4 Q8 FP16	143 tok/s 106 tok/s 67 tok/s	~830 ms ~832 ms ~838 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	133 tok/s 97 tok/s 60 tok/s	~953 ms ~956 ms ~962 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	133 tok/s 97 tok/s 60 tok/s	~953 ms ~956 ms ~962 ms
-	SVD Image-to-Video XT	1.5B	INT4 Q8 FP16	133 tok/s 97 tok/s 60 tok/s	~953 ms ~956 ms ~962 ms
-	SVD Image-to-Video XT 1.1	1.5B	INT4 Q8 FP16	133 tok/s 97 tok/s 60 tok/s	~953 ms ~956 ms ~962 ms
-	LFM3 1.7B	1.7B	INT4 Q8 FP16	124 tok/s 89 tok/s 54 tok/s	~1.1s ~1.1s ~1.1s
-	LFM3 1.7B Base	1.7B	INT4 Q8 FP16	124 tok/s 89 tok/s 54 tok/s	~1.1s ~1.1s ~1.1s
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	124 tok/s 89 tok/s 54 tok/s	~1.1s ~1.1s ~1.1s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	120 tok/s 85 tok/s 52 tok/s	~1.1s ~1.1s ~1.1s
-	CogVideoX 2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	LFM3 Audio 2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	LFM3 Omni 2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	LFM3 VL 2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	Liquid D1 2B	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	LTX-Video 0.9.1	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	Moondream 3.1	9B (2B active)	INT4	113 tok/s	~1.4s
-	Stable Diffusion 3 Medium	2B	INT4 Q8 FP16	113 tok/s 79 tok/s 48 tok/s	~1.3s ~1.3s ~1.3s
-	Kandinsky 3.0	2.1B	INT4 Q8 FP16	110 tok/s 77 tok/s 46 tok/s	~1.3s ~1.3s ~1.3s
-	LTX-2	2.2B	INT4 Q8 FP16	107 tok/s 74 tok/s 44 tok/s	~1.4s ~1.4s ~1.4s
-	SDXL 1.0 Refiner	2.31B	INT4 Q8 FP16	105 tok/s 72 tok/s 42 tok/s	~1.4s ~1.4s ~1.4s
-	Stable Diffusion 3.5 Medium	2.5B	INT4 Q8 FP16	100 tok/s 68 tok/s 40 tok/s	~1.5s ~1.5s ~1.6s
-	SDXL 1.0 Base	2.56B	INT4 Q8 FP16	98 tok/s 67 tok/s 37 tok/s	~1.6s ~1.6s ~1.7s
-	SDXL Turbo	2.56B	INT4 Q8 FP16	98 tok/s 67 tok/s 37 tok/s	~1.6s ~1.6s ~1.7s
-	Falcon-3B	3B	INT4 Q8 FP16	95 tok/s 61 tok/s 33 tok/s	~1.8s ~1.8s ~2.0s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	93 tok/s 77 tok/s 54 tok/s	~834 ms ~836 ms ~841 ms
-	Phi-1	1.3B	INT4 Q8 FP16	93 tok/s 77 tok/s 54 tok/s	~834 ms ~836 ms ~841 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	93 tok/s 77 tok/s 54 tok/s	~834 ms ~836 ms ~841 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	91 tok/s 60 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	Falcon3-3B	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	Liquid D1 3B	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	89 tok/s 59 tok/s 32 tok/s	~1.8s ~1.8s ~2.0s
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	82 tok/s	~2.1s
-	FLUX.2 Klein 4B	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
-	FrogNano 4B	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
-	Mage-VL Cowork	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
-	Spark X2.5 4B	4B	INT4 Q8	74 tok/s 47 tok/s	~2.4s ~2.4s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	70 tok/s 57 tok/s 37 tok/s	~1.3s ~1.3s ~1.4s
-	CogVideoX 1.5 5B	5B	INT4 Q8	63 tok/s 37 tok/s	~2.9s ~3.2s
-	CogVideoX 5B	5B	INT4 Q8	63 tok/s 37 tok/s	~2.9s ~3.2s
-	Wan 2.2 TI2V 5B	5B	INT4 Q8	63 tok/s 37 tok/s	~2.9s ~3.2s
-	Phi-2	2.7B	INT4 Q8 FP16	62 tok/s 48 tok/s 28 tok/s	~1.7s ~1.7s ~2.0s
-	MaLLaM-3B	3B	INT4 Q8	59 tok/s 45 tok/s	~1.8s ~1.8s
-	SmolLM3 3B	3B	INT4 Q8	59 tok/s 45 tok/s	~1.8s ~1.8s
-	Falcon-7B	7B	INT4 Q8	51 tok/s 27 tok/s	~4.0s ~4.8s
-	AuraFlow v0.2	6.8B	INT4 Q8	49 tok/s 27 tok/s	~3.9s ~4.6s
-	AuraFlow v0.3	6.8B	INT4 Q8	49 tok/s 27 tok/s	~3.9s ~4.6s
-	Falcon3-7B	7B	INT4	48 tok/s	~4.0s
-	Hy-MT2 7B	7B	INT4	48 tok/s	~4.0s
-	Mistral-7B-v0.1	7.3B	INT4	47 tok/s	~4.2s
-	Youtu Parsing Omni	8B	INT4	44 tok/s	~4.6s
-	Stable Diffusion 3.5 Large	8.1B	INT4	43 tok/s	~4.7s
-	Stable Diffusion 3.5 Large Turbo	8.1B	INT4	43 tok/s	~4.7s
-	FLUX.2 Klein 9B	9B	INT4	39 tok/s	~5.2s
-	LensVLM 9B	9B	INT4	39 tok/s	~5.2s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	39 tok/s	~5.2s
-	ChatGLM3-6B-32K	6B	INT4	36 tok/s	~3.8s
-	Yi-6B	6B	INT4	36 tok/s	~3.8s
-	Falcon3-10B	10B	INT4	34 tok/s	~6.2s
-	Mochi 1 Preview	10B	INT4	34 tok/s	~6.2s
-	Falcon2-11B	11B	INT4	33 tok/s	~6.8s
-	Hunyuan Lite	7B	INT4	32 tok/s	~4.4s
-	OLMo 3 7B Base	7B	INT4	32 tok/s	~4.4s
-	SEA-LION-7B	7.1B	INT4	32 tok/s	~4.5s
-	SEA-LION-7B-Instruct	7.1B	INT4	32 tok/s	~4.5s
-	FLUX.1 Canny Dev	11.9B	INT4	28 tok/s	~8.0s
-	FLUX.1 Depth Dev	11.9B	INT4	28 tok/s	~8.0s
-	FLUX.1 Dev	11.9B	INT4	28 tok/s	~8.0s
-	FLUX.1 Fill Dev	11.9B	INT4	28 tok/s	~8.0s
-	FLUX.1 Schnell	11.9B	INT4	28 tok/s	~8.0s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	28 tok/s	~5.4s
-	Typhoon-2-8B	8B	INT4	28 tok/s	~5.4s
-	HunyuanVideo 1.5 720P T2V	13B	INT4	26 tok/s	~8.7s

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