What are the best models to run on Radeon PRO W6600 (8GB)?

Based on model rankings and 8 GB usable VRAM: Top Overall: Qwen3.5-9B (INT4) ~33 tok/s. Top Coding: K2 Horizon 7B (INT4) ~41 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~37 tok/s. Maximum Model Size: Mellum 2.1 12B A2.5B Thinking (INT4) ~74 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 Radeon PRO W6600 (8GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#121	Qwen3.5-9B	9B	INT4	33 tok/s	~5.0s
#127	K2 Horizon 7B	7B	INT4	41 tok/s	~3.9s
#136	Qwen3.5-4B	4B	INT4 Q8	65 tok/s 40 tok/s	~2.3s ~2.3s
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	10 tok/s 6 tok/s	~15.7s ~17.1s
#152	Granite 4.2 8B	8B	INT4	37 tok/s	~4.4s
#157	Gemma 4 E4B	8B	INT4	37 tok/s	~4.4s
#180	Gemma 3 12B	12B	INT4	23 tok/s	~7.8s
#185	Ministral 3 8B	8B	INT4	23 tok/s	~5.2s
#187	OLMo 3 7B Think	7B	INT4	27 tok/s	~4.2s
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~1.2s ~1.2s ~1.2s
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~1.2s ~1.2s ~1.2s
#193	Gemma 4 E2B	5.1B	INT4 Q8	54 tok/s 31 tok/s	~2.9s ~3.1s
#195	Ministral 3 3B	3B	INT4 Q8	51 tok/s 38 tok/s	~1.8s ~1.8s
#196	DeepSeek-R1 8B	8B	INT4	38 tok/s	~4.4s
#198	Gemma 2 9B	9B	INT4	33 tok/s	~5.0s
#201	Qwen3-4B	4B	INT4 Q8	65 tok/s 40 tok/s	~2.3s ~2.3s
#203	Gemma 3 4B	4B	INT4 Q8	65 tok/s 40 tok/s	~2.3s ~2.3s
#204	DeepSeek-R1 7B	7B	INT4 Q8	42 tok/s 22 tok/s	~3.9s ~4.6s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	186 tok/s 137 tok/s 87 tok/s	~493 ms ~495 ms ~500 ms
#207	Llama 3.1 8B	8B	INT4	37 tok/s	~4.4s
#207	Qwen3-8B	8B	INT4	37 tok/s	~4.4s
#212	Llama 3 8B	8B	INT4	37 tok/s	~4.4s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	119 tok/s 81 tok/s 47 tok/s	~1.0s ~1.0s ~1.0s
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	235 tok/s 207 tok/s 160 tok/s	~165 ms ~166 ms ~167 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	214 tok/s 164 tok/s 108 tok/s	~365 ms ~367 ms ~370 ms
#217	Phi-4-Mini	3.8B	INT4 Q8	68 tok/s 42 tok/s	~2.2s ~2.2s
#218	Qwen2.5-7B	7B	INT4	41 tok/s	~3.9s
#220	Ministral-8B-2410	8B	INT4	37 tok/s	~4.4s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	129 tok/s 89 tok/s 52 tok/s	~917 ms ~920 ms ~928 ms
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
#223	Phi-3-mini	3.8B	INT4 Q8	68 tok/s 42 tok/s	~2.2s ~2.2s
#224	Gemma 3n E2B IT	6B (2B active)	INT4	48 tok/s	~1.4s
#225	Gemma 2 2B	2B	INT4 Q8 FP16	107 tok/s 71 tok/s 41 tok/s	~1.2s ~1.2s ~1.2s
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4	40 tok/s	~4.0s
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	163 tok/s 118 tok/s 73 tok/s	~627 ms ~629 ms ~634 ms
#230	OLMo 3 7B Instruct	7B	INT4	27 tok/s	~4.2s
#231	Qwen2-7B	7B	INT4	41 tok/s	~3.9s
#232	Phi-3-small	7B	INT4	41 tok/s	~3.9s
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4	40 tok/s	~4.0s
#234	Gemma 3 1B	1B	INT4 Q8 FP16	163 tok/s 118 tok/s 73 tok/s	~627 ms ~629 ms ~634 ms
#236	Gemma 1 7B	7B	INT4	27 tok/s	~4.2s
#237	Gemma 1 2B	2B	INT4 Q8 FP16	116 tok/s 75 tok/s 42 tok/s	~1.2s ~1.2s ~1.2s
#238	ChatGLM3-6B	6B	INT4	31 tok/s	~3.7s
#239	ChatGLM2-6B	6B	INT4	31 tok/s	~3.7s
#240	ChatGLM-6B	6B	INT4	31 tok/s	~3.7s
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	272 tok/s 228 tok/s 168 tok/s	~183 ms ~184 ms ~185 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	231 tok/s 181 tok/s 123 tok/s	~303 ms ~304 ms ~307 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	231 tok/s 181 tok/s 123 tok/s	~303 ms ~304 ms ~307 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	224 tok/s 196 tok/s 150 tok/s	~184 ms ~184 ms ~186 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	224 tok/s 196 tok/s 150 tok/s	~184 ms ~184 ms ~186 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	214 tok/s 164 tok/s 108 tok/s	~365 ms ~367 ms ~370 ms
-	Falcon-1B	1B	INT4 Q8 FP16	176 tok/s 124 tok/s 75 tok/s	~626 ms ~629 ms ~634 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	163 tok/s 118 tok/s 73 tok/s	~627 ms ~629 ms ~634 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	163 tok/s 118 tok/s 73 tok/s	~627 ms ~629 ms ~634 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	129 tok/s 89 tok/s 52 tok/s	~917 ms ~920 ms ~928 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	129 tok/s 89 tok/s 52 tok/s	~917 ms ~920 ms ~928 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	119 tok/s 81 tok/s 47 tok/s	~1.0s ~1.0s ~1.0s
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	115 tok/s 77 tok/s 45 tok/s	~1.1s ~1.1s ~1.1s
-	Falcon-3B	3B	INT4 Q8 FP16	87 tok/s 54 tok/s 28 tok/s	~1.7s ~1.8s ~1.9s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	85 tok/s 69 tok/s 47 tok/s	~803 ms ~806 ms ~813 ms
-	Phi-1	1.3B	INT4 Q8 FP16	85 tok/s 69 tok/s 47 tok/s	~803 ms ~806 ms ~813 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	85 tok/s 69 tok/s 47 tok/s	~803 ms ~806 ms ~813 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	82 tok/s 52 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	Falcon3-3B	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	Liquid D1 3B	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	81 tok/s 51 tok/s 27 tok/s	~1.7s ~1.8s ~1.9s
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	74 tok/s	~2.0s
-	FrogNano 4B	4B	INT4 Q8	65 tok/s 40 tok/s	~2.3s ~2.3s
-	Spark X2.5 4B	4B	INT4 Q8	65 tok/s 40 tok/s	~2.3s ~2.3s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	62 tok/s 49 tok/s 31 tok/s	~1.2s ~1.2s ~1.3s
-	Phi-2	2.7B	INT4 Q8 FP16	54 tok/s 41 tok/s 23 tok/s	~1.6s ~1.6s ~1.9s
-	MaLLaM-3B	3B	INT4 Q8	51 tok/s 38 tok/s	~1.8s ~1.8s
-	SmolLM3 3B	3B	INT4 Q8	51 tok/s 38 tok/s	~1.8s ~1.8s
-	Falcon-7B	7B	INT4 Q8	43 tok/s 23 tok/s	~3.9s ~4.6s
-	Falcon3-7B	7B	INT4	41 tok/s	~3.9s
-	Hy-MT2 7B	7B	INT4	41 tok/s	~3.9s
-	Mistral-7B-v0.1	7.3B	INT4	40 tok/s	~4.0s
-	Youtu Parsing Omni	8B	INT4	38 tok/s	~4.4s
-	LensVLM 9B	9B	INT4	33 tok/s	~5.0s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	33 tok/s	~5.0s
-	ChatGLM3-6B-32K	6B	INT4	31 tok/s	~3.7s
-	Yi-6B	6B	INT4	31 tok/s	~3.7s
-	Falcon3-10B	10B	INT4	29 tok/s	~6.0s
-	Falcon2-11B	11B	INT4	27 tok/s	~6.6s
-	Hunyuan Lite	7B	INT4	27 tok/s	~4.2s
-	OLMo 3 7B Base	7B	INT4	27 tok/s	~4.2s
-	SEA-LION-7B	7.1B	INT4	27 tok/s	~4.3s
-	SEA-LION-7B-Instruct	7.1B	INT4	27 tok/s	~4.3s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	23 tok/s	~5.2s
-	Typhoon-2-8B	8B	INT4	23 tok/s	~5.2s

Radeon PRO W6600 (8GB)

VRAM Scaling by Context Length

Best Models to Run on Radeon PRO W6600 (8GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions