What are the best models to run on Radeon PRO W7500 (8GB)?

Based on model rankings and 8 GB usable VRAM: Top Overall: K2 Horizon 7B (INT4) ~33 tok/s. Top Coding: Qwen3.5-9B (INT4) ~27 tok/s. Optimal Daily Driver: Granite 4.2 8B (INT4) ~30 tok/s. Maximum Model Size: Mellum 2.1 12B A2.5B Thinking (INT4) ~61 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 W7500 (8GB) (October 2026)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#118	K2 Horizon 7B	7B	INT4	33 tok/s	~3.3s
#132	Qwen3.5-4B	4B	INT4 Q8	53 tok/s 32 tok/s	~1.9s ~1.9s
#133	Qwen3.5-9B	9B	INT4	27 tok/s	~4.2s
#141	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8	8 tok/s 4 tok/s	~13.1s ~14.3s
#149	Granite 4.2 8B	8B	INT4	30 tok/s	~3.7s
#155	Gemma 4 E4B	8B	INT4	30 tok/s	~3.7s
#177	Gemma 3 12B	12B	INT4	18 tok/s	~6.5s
#183	Ministral 3 8B	8B	INT4	19 tok/s	~4.4s
#184	Qwen3.5-2B	2B	INT4 Q8 FP16	89 tok/s 58 tok/s 33 tok/s	~1.0s ~1.0s ~1.0s
#186	MiniCPM5-2B	2B	INT4 Q8 FP16	89 tok/s 58 tok/s 33 tok/s	~1.0s ~1.0s ~1.0s
#188	OLMo 3 7B Think	7B	INT4	22 tok/s	~3.5s
#190	Gemma 4 E2B	5.1B	INT4 Q8	44 tok/s 25 tok/s	~2.4s ~2.6s
#192	Ministral 3 3B	3B	INT4 Q8	41 tok/s 31 tok/s	~1.5s ~1.5s
#194	DeepSeek-R1 8B	8B	INT4	30 tok/s	~3.7s
#195	Gemma 2 9B	9B	INT4	27 tok/s	~4.2s
#197	Gemma 3 4B	4B	INT4 Q8	53 tok/s 32 tok/s	~1.9s ~1.9s
#198	Qwen2.5-7B	7B	INT4	33 tok/s	~3.3s
#199	DeepSeek-R1 7B	7B	INT4 Q8	34 tok/s 18 tok/s	~3.3s ~3.8s
#202	Qwen3-8B	8B	INT4	30 tok/s	~3.7s
#203	Llama 3.1 8B	8B	INT4	30 tok/s	~3.7s
#204	Qwen3-4B	4B	INT4 Q8	53 tok/s 32 tok/s	~1.9s ~1.9s
#209	Llama 3 8B	8B	INT4	30 tok/s	~3.7s
#210	Phi-4-Mini	3.8B	INT4 Q8	56 tok/s 34 tok/s	~1.8s ~1.8s
#211	Gemma 3 270M	0.27B	INT4 Q8 FP16	214 tok/s 185 tok/s 138 tok/s	~140 ms ~140 ms ~142 ms
#212	Qwen3-1.7B	1.7B	INT4 Q8 FP16	101 tok/s 66 tok/s 38 tok/s	~867 ms ~872 ms ~884 ms
#213	Qwen3-0.6B	0.6B	INT4 Q8 FP16	192 tok/s 142 tok/s 91 tok/s	~307 ms ~308 ms ~312 ms
#214	Ministral-8B-2410	8B	INT4	30 tok/s	~3.7s
#216	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	110 tok/s 73 tok/s 42 tok/s	~768 ms ~772 ms ~782 ms
#217	Llama 3.2 3B	3B	INT4 Q8 FP16	67 tok/s 42 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
#218	Phi-3-mini	3.8B	INT4 Q8	56 tok/s 34 tok/s	~1.8s ~1.8s
#219	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	163 tok/s 117 tok/s 72 tok/s	~414 ms ~416 ms ~421 ms
#220	Gemma 3n E2B IT	6B (2B active)	INT4	39 tok/s	~1.2s
#221	Gemma 2 2B	2B	INT4 Q8 FP16	89 tok/s 58 tok/s 33 tok/s	~1.0s ~1.0s ~1.0s
#223	Mistral-7B-Instruct-v0.1	7.3B	INT4	32 tok/s	~3.4s
#224	Llama 3.2 1B	1B	INT4 Q8 FP16	142 tok/s 99 tok/s 60 tok/s	~525 ms ~528 ms ~535 ms
#226	OLMo 3 7B Instruct	7B	INT4	22 tok/s	~3.5s
#227	Qwen2-7B	7B	INT4	33 tok/s	~3.3s
#228	Phi-3-small	7B	INT4	33 tok/s	~3.3s
#229	Gemma 3 1B	1B	INT4 Q8 FP16	142 tok/s 99 tok/s 60 tok/s	~525 ms ~528 ms ~535 ms
#230	Mistral-7B-Instruct-v0.2	7.3B	INT4	32 tok/s	~3.4s
#232	Gemma 1 7B	7B	INT4	22 tok/s	~3.5s
#233	Gemma 1 2B	2B	INT4 Q8 FP16	97 tok/s 61 tok/s 34 tok/s	~1.0s ~1.0s ~1.0s
#234	ChatGLM3-6B	6B	INT4	24 tok/s	~3.1s
#235	ChatGLM2-6B	6B	INT4	24 tok/s	~3.1s
#236	ChatGLM-6B	6B	INT4	24 tok/s	~3.1s
Unranked Models (Sorted by Speed)
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	209 tok/s 159 tok/s 104 tok/s	~255 ms ~256 ms ~260 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	209 tok/s 159 tok/s 104 tok/s	~255 ms ~256 ms ~260 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	202 tok/s 173 tok/s 129 tok/s	~155 ms ~156 ms ~158 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	202 tok/s 173 tok/s 129 tok/s	~155 ms ~156 ms ~158 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	192 tok/s 142 tok/s 91 tok/s	~307 ms ~308 ms ~312 ms
-	Falcon-1B	1B	INT4 Q8 FP16	154 tok/s 105 tok/s 62 tok/s	~525 ms ~528 ms ~534 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	142 tok/s 99 tok/s 60 tok/s	~525 ms ~528 ms ~535 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	110 tok/s 73 tok/s 42 tok/s	~768 ms ~772 ms ~782 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	110 tok/s 73 tok/s 42 tok/s	~768 ms ~772 ms ~782 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	101 tok/s 66 tok/s 38 tok/s	~867 ms ~872 ms ~884 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	97 tok/s 64 tok/s 36 tok/s	~913 ms ~918 ms ~931 ms
-	Falcon-3B	3B	INT4 Q8 FP16	72 tok/s 43 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	70 tok/s 57 tok/s 38 tok/s	~675 ms ~678 ms ~687 ms
-	Phi-1	1.3B	INT4 Q8 FP16	70 tok/s 57 tok/s 38 tok/s	~675 ms ~678 ms ~687 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	70 tok/s 57 tok/s 38 tok/s	~675 ms ~678 ms ~687 ms
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	68 tok/s 42 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
-	Falcon3-3B	3B	INT4 Q8 FP16	67 tok/s 42 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	67 tok/s 42 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	67 tok/s 42 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	67 tok/s 42 tok/s 22 tok/s	~1.5s ~1.5s ~1.6s
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	61 tok/s	~1.7s
-	FrogNano 4B	4B	INT4 Q8	53 tok/s 32 tok/s	~1.9s ~1.9s
-	Spark X2.5 4B	4B	INT4 Q8	53 tok/s 32 tok/s	~1.9s ~1.9s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	51 tok/s 40 tok/s 25 tok/s	~1.0s ~1.0s ~1.1s
-	Phi-2	2.7B	INT4 Q8 FP16	44 tok/s 33 tok/s 19 tok/s	~1.3s ~1.3s ~1.6s
-	MaLLaM-3B	3B	INT4 Q8	41 tok/s 31 tok/s	~1.5s ~1.5s
-	SmolLM3 3B	3B	INT4 Q8	41 tok/s 31 tok/s	~1.5s ~1.5s
-	Falcon-7B	7B	INT4 Q8	35 tok/s 18 tok/s	~3.3s ~3.8s
-	Falcon3-7B	7B	INT4	33 tok/s	~3.3s
-	Hy-MT2 7B	7B	INT4	33 tok/s	~3.3s
-	Mistral-7B-v0.1	7.3B	INT4	32 tok/s	~3.4s
-	LensVLM 9B	9B	INT4	27 tok/s	~4.2s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4	27 tok/s	~4.2s
-	ChatGLM3-6B-32K	6B	INT4	24 tok/s	~3.1s
-	Yi-6B	6B	INT4	24 tok/s	~3.1s
-	Falcon3-10B	10B	INT4	23 tok/s	~5.0s
-	Falcon2-11B	11B	INT4	22 tok/s	~5.5s
-	Hunyuan Lite	7B	INT4	22 tok/s	~3.5s
-	OLMo 3 7B Base	7B	INT4	22 tok/s	~3.5s
-	SEA-LION-7B	7.1B	INT4	22 tok/s	~3.6s
-	SEA-LION-7B-Instruct	7.1B	INT4	22 tok/s	~3.6s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4	19 tok/s	~4.4s
-	Typhoon-2-8B	8B	INT4	19 tok/s	~4.4s

Radeon PRO W7500 (8GB)

VRAM Scaling by Context Length

Best Models to Run on Radeon PRO W7500 (8GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions