What are the best models to run on Radeon PRO W7600 (8GB)?

Based on model rankings and 8 GB usable VRAM: Top Overall: K2 Horizon 7B (INT4) ~51 tok/s. Top Coding: Qwen3.5-9B (INT4) ~42 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~46 tok/s. Maximum Model Size: Mellum 2.1 12B A2.5B Thinking (INT4) ~90 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 Radeon PRO W7600 (8GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#118	K2 Horizon 7B	7B	INT4	51 tok/s	~2.0s
#132	Qwen3.5-4B	4B	INT4 Q8	80 tok/s 50 tok/s	~1.1s ~1.2s
#133	Qwen3.5-9B	9B	INT4	42 tok/s	~2.5s
#141	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	13 tok/s 7 tok/s	~7.9s ~8.6s
#149	Granite 4.2 8B	8B	INT4	46 tok/s	~2.2s
#155	Gemma 4 E4B	8B	INT4	46 tok/s	~2.2s
#177	Gemma 3 12B	12B	INT4	29 tok/s	~3.9s
#183	Ministral 3 8B	8B	INT4	29 tok/s	~2.6s
#184	Qwen3.5-2B	2B	INT4 Q8 FP16	127 tok/s 87 tok/s 51 tok/s	~612 ms ~615 ms ~624 ms
#186	MiniCPM5-2B	2B	INT4 Q8 FP16	127 tok/s 87 tok/s 51 tok/s	~612 ms ~615 ms ~624 ms
#188	OLMo 3 7B Think	7B	INT4	34 tok/s	~2.1s
#190	Gemma 4 E2B	5.1B	INT4 Q8	67 tok/s 39 tok/s	~1.4s ~1.6s
#192	Ministral 3 3B	3B	INT4 Q8	63 tok/s 48 tok/s	~885 ms ~890 ms
#194	DeepSeek-R1 8B	8B	INT4	47 tok/s	~2.2s
#195	Gemma 2 9B	9B	INT4	42 tok/s	~2.5s
#197	Gemma 3 4B	4B	INT4 Q8	80 tok/s 50 tok/s	~1.1s ~1.2s
#198	Qwen2.5-7B	7B	INT4	51 tok/s	~2.0s
#199	DeepSeek-R1 7B	7B	INT4 Q8	53 tok/s 28 tok/s	~2.0s ~2.3s
#202	Qwen3-8B	8B	INT4	46 tok/s	~2.2s
#203	Llama 3.1 8B	8B	INT4	46 tok/s	~2.2s
#204	Qwen3-4B	4B	INT4 Q8	80 tok/s 50 tok/s	~1.1s ~1.2s
#209	Llama 3 8B	8B	INT4	46 tok/s	~2.2s
#210	Phi-4-Mini	3.8B	INT4 Q8	83 tok/s 53 tok/s	~1.1s ~1.1s
#211	Gemma 3 270M	0.27B	INT4 Q8 FP16	256 tok/s 229 tok/s 183 tok/s	~87 ms ~87 ms ~89 ms
#212	Qwen3-1.7B	1.7B	INT4 Q8 FP16	140 tok/s 98 tok/s 58 tok/s	~524 ms ~527 ms ~534 ms
#213	Qwen3-0.6B	0.6B	INT4 Q8 FP16	236 tok/s 187 tok/s 128 tok/s	~187 ms ~188 ms ~191 ms
#214	Ministral-8B-2410	8B	INT4	46 tok/s	~2.2s
#216	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	151 tok/s 107 tok/s 65 tok/s	~464 ms ~467 ms ~473 ms
#217	Llama 3.2 3B	3B	INT4 Q8 FP16	98 tok/s 64 tok/s 34 tok/s	~879 ms ~885 ms ~973 ms
#218	Phi-3-mini	3.8B	INT4 Q8	83 tok/s 53 tok/s	~1.1s ~1.1s
#219	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	209 tok/s 160 tok/s 105 tok/s	~252 ms ~253 ms ~256 ms
#220	Gemma 3n E2B IT	6B (2B active)	INT4	60 tok/s	~717 ms
#221	Gemma 2 2B	2B	INT4 Q8 FP16	127 tok/s 87 tok/s 51 tok/s	~612 ms ~615 ms ~624 ms
#223	Mistral-7B-Instruct-v0.1	7.3B	INT4	50 tok/s	~2.0s
#224	Llama 3.2 1B	1B	INT4 Q8 FP16	186 tok/s 139 tok/s 89 tok/s	~318 ms ~320 ms ~324 ms
#226	OLMo 3 7B Instruct	7B	INT4	34 tok/s	~2.1s
#227	Qwen2-7B	7B	INT4	51 tok/s	~2.0s
#228	Phi-3-small	7B	INT4	51 tok/s	~2.0s
#229	Gemma 3 1B	1B	INT4 Q8 FP16	186 tok/s 139 tok/s 89 tok/s	~318 ms ~320 ms ~324 ms
#230	Mistral-7B-Instruct-v0.2	7.3B	INT4	50 tok/s	~2.0s
#232	Gemma 1 7B	7B	INT4	34 tok/s	~2.1s
#233	Gemma 1 2B	2B	INT4 Q8 FP16	137 tok/s 91 tok/s 52 tok/s	~611 ms ~615 ms ~623 ms
#234	ChatGLM3-6B	6B	INT4	38 tok/s	~1.8s
#235	ChatGLM2-6B	6B	INT4	38 tok/s	~1.8s
#236	ChatGLM-6B	6B	INT4	38 tok/s	~1.8s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	252 tok/s 204 tok/s 144 tok/s	~156 ms ~157 ms ~159 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	252 tok/s 204 tok/s 144 tok/s	~156 ms ~157 ms ~159 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	245 tok/s 218 tok/s 172 tok/s	~96 ms ~97 ms ~98 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	245 tok/s 218 tok/s 172 tok/s	~96 ms ~97 ms ~98 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	236 tok/s 187 tok/s 128 tok/s	~187 ms ~188 ms ~191 ms
-	Falcon-1B	1B	INT4 Q8 FP16	199 tok/s 145 tok/s 91 tok/s	~318 ms ~320 ms ~324 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	186 tok/s 139 tok/s 89 tok/s	~318 ms ~320 ms ~324 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	151 tok/s 107 tok/s 65 tok/s	~464 ms ~467 ms ~473 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	151 tok/s 107 tok/s 65 tok/s	~464 ms ~467 ms ~473 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	140 tok/s 98 tok/s 58 tok/s	~524 ms ~527 ms ~534 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	136 tok/s 94 tok/s 55 tok/s	~551 ms ~555 ms ~562 ms
-	Falcon-3B	3B	INT4 Q8 FP16	105 tok/s 66 tok/s 35 tok/s	~879 ms ~884 ms ~973 ms
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	103 tok/s 84 tok/s 58 tok/s	~408 ms ~410 ms ~416 ms
-	Phi-1	1.3B	INT4 Q8 FP16	103 tok/s 84 tok/s 58 tok/s	~408 ms ~410 ms ~416 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	103 tok/s 84 tok/s 58 tok/s	~408 ms ~410 ms ~416 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	100 tok/s 64 tok/s 34 tok/s	~879 ms ~885 ms ~973 ms
-	Falcon3-3B	3B	INT4 Q8 FP16	98 tok/s 64 tok/s 34 tok/s	~879 ms ~885 ms ~973 ms
-	LFM2-VL 3B	3B	INT4 Q8 FP16	98 tok/s 64 tok/s 34 tok/s	~879 ms ~885 ms ~973 ms
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	98 tok/s 64 tok/s 34 tok/s	~879 ms ~885 ms ~973 ms
-	Qwen2.5-3B	3B	INT4 Q8 FP16	98 tok/s 64 tok/s 34 tok/s	~879 ms ~885 ms ~973 ms
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	90 tok/s	~1.0s
-	FrogNano 4B	4B	INT4 Q8	80 tok/s 50 tok/s	~1.1s ~1.2s
-	Spark X2.5 4B	4B	INT4 Q8	80 tok/s 50 tok/s	~1.1s ~1.2s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	76 tok/s 61 tok/s 39 tok/s	~617 ms ~620 ms ~682 ms
-	Phi-2	2.7B	INT4 Q8 FP16	66 tok/s 51 tok/s 29 tok/s	~805 ms ~809 ms ~962 ms
-	MaLLaM-3B	3B	INT4 Q8	63 tok/s 48 tok/s	~885 ms ~890 ms
-	SmolLM3 3B	3B	INT4 Q8	63 tok/s 48 tok/s	~885 ms ~890 ms
-	Falcon-7B	7B	INT4 Q8	54 tok/s 29 tok/s	~2.0s ~2.3s
-	Falcon3-7B	7B	INT4	51 tok/s	~2.0s
-	Hy-MT2 7B	7B	INT4	51 tok/s	~2.0s
-	Mistral-7B-v0.1	7.3B	INT4	50 tok/s	~2.0s
-	LensVLM 9B	9B	INT4	42 tok/s	~2.5s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	42 tok/s	~2.5s
-	ChatGLM3-6B-32K	6B	INT4	38 tok/s	~1.8s
-	Yi-6B	6B	INT4	38 tok/s	~1.8s
-	Falcon3-10B	10B	INT4	36 tok/s	~3.0s
-	Falcon2-11B	11B	INT4	35 tok/s	~3.3s
-	Hunyuan Lite	7B	INT4	34 tok/s	~2.1s
-	OLMo 3 7B Base	7B	INT4	34 tok/s	~2.1s
-	SEA-LION-7B	7.1B	INT4	34 tok/s	~2.2s
-	SEA-LION-7B-Instruct	7.1B	INT4	34 tok/s	~2.2s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	29 tok/s	~2.6s
-	Typhoon-2-8B	8B	INT4	29 tok/s	~2.6s

Radeon PRO W7600 (8GB)

VRAM Scaling by Context Length

Best Models to Run on Radeon PRO W7600 (8GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions