What are the best models to run on Arc B580 (12GB)?

Based on model rankings and 12 GB usable VRAM: Top Overall: Phi-4 Reasoning Plus (INT4) ~28 tok/s. Top Coding: K2 Horizon 7B (INT4) ~75 tok/s. Optimal Daily Driver: Qwen3.5-9B (INT4) ~62 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 Arc B580 (12GB)

Rank	Model	Parameters	Precision	TPS (Approx.)	TTFT
#102	Phi-4 Reasoning Plus	14B	INT4	28 tok/s	~6.6s
#121	Qwen3.5-9B	9B	INT4 Q8	62 tok/s 36 tok/s	~3.6s ~4.0s
#127	K2 Horizon 7B	7B	INT4 Q8	75 tok/s 47 tok/s	~2.8s ~2.9s
#130	Gemma 4 12B	11.95B	INT4	34 tok/s	~5.2s
#136	Qwen3.5-4B	4B	INT4 Q8 FP16	113 tok/s 74 tok/s 42 tok/s	~1.7s ~1.7s ~1.7s
#140	GLM-4V	9B	INT4	45 tok/s	~3.6s
#145	NVIDIA Nemotron 3 Nano 30B-A3B	3.5B (30B active)	INT4 Q8 FP16	20 tok/s 12 tok/s 6 tok/s	~11.5s ~11.5s ~12.6s
#152	Granite 4.2 8B	8B	INT4 Q8	68 tok/s 39 tok/s	~3.2s ~3.5s
#157	Gemma 4 E4B	8B	INT4 Q8	68 tok/s 39 tok/s	~3.2s ~3.5s
#164	Phi-4	14B	INT4	42 tok/s	~5.6s
#176	DeepSeek-R1 14B	14B	INT4	44 tok/s	~5.6s
#177	Ministral 3 14B	14B	INT4	28 tok/s	~6.6s
#180	Gemma 3 12B	12B	INT4	48 tok/s	~4.8s
#185	Ministral 3 8B	8B	INT4 Q8	49 tok/s 31 tok/s	~3.2s ~3.8s
#187	OLMo 3 7B Think	7B	INT4 Q8	54 tok/s 36 tok/s	~2.9s ~3.1s
#188	Qwen3.5-2B	2B	INT4 Q8 FP16	167 tok/s 121 tok/s 74 tok/s	~887 ms ~889 ms ~894 ms
#189	MiniCPM5-2B	2B	INT4 Q8 FP16	167 tok/s 121 tok/s 74 tok/s	~887 ms ~889 ms ~894 ms
#190	Qwen2.5-14B	14B	INT4	42 tok/s	~5.6s
#193	Gemma 4 E2B	5.1B	INT4 Q8 FP16	95 tok/s 61 tok/s 30 tok/s	~2.1s ~2.1s ~2.5s
#195	Ministral 3 3B	3B	INT4 Q8 FP16	91 tok/s 70 tok/s 45 tok/s	~1.3s ~1.3s ~1.3s
#196	DeepSeek-R1 8B	8B	INT4 Q8	70 tok/s 42 tok/s	~3.2s ~3.2s
#198	Gemma 2 9B	9B	INT4 Q8	62 tok/s 36 tok/s	~3.6s ~4.0s
#201	Qwen3-4B	4B	INT4 Q8 FP16	113 tok/s 74 tok/s 42 tok/s	~1.7s ~1.7s ~1.7s
#202	Qwen3-14B	14B	INT4	42 tok/s	~5.6s
#203	Gemma 3 4B	4B	INT4 Q8 FP16	113 tok/s 74 tok/s 42 tok/s	~1.7s ~1.7s ~1.7s
#204	DeepSeek-R1 7B	7B	INT4 Q8	77 tok/s 47 tok/s	~2.8s ~2.9s
#206	Qwen3.5-0.8B	0.8B	INT4 Q8 FP16	249 tok/s 202 tok/s 143 tok/s	~362 ms ~363 ms ~365 ms
#207	Llama 3.1 8B	8B	INT4 Q8	68 tok/s 39 tok/s	~3.2s ~3.5s
#207	Qwen3-8B	8B	INT4 Q8	68 tok/s 39 tok/s	~3.2s ~3.5s
#212	Llama 3 8B	8B	INT4 Q8	68 tok/s 39 tok/s	~3.2s ~3.5s
#213	Qwen3-1.7B	1.7B	INT4 Q8 FP16	182 tok/s 134 tok/s 84 tok/s	~759 ms ~761 ms ~765 ms
#214	Gemma 3 270M	0.27B	INT4 Q8 FP16	289 tok/s 267 tok/s 225 tok/s	~123 ms ~123 ms ~123 ms
#216	Qwen3-0.6B	0.6B	INT4 Q8 FP16	272 tok/s 229 tok/s 169 tok/s	~269 ms ~270 ms ~271 ms
#217	Phi-4-Mini	3.8B	INT4 Q8 FP16	116 tok/s 77 tok/s 44 tok/s	~1.6s ~1.6s ~1.6s
#218	Qwen2.5-7B	7B	INT4 Q8	75 tok/s 47 tok/s	~2.8s ~2.9s
#220	Ministral-8B-2410	8B	INT4 Q8	68 tok/s 39 tok/s	~3.2s ~3.5s
#221	DeepSeek-R1 1.5B	1.5B	INT4 Q8 FP16	193 tok/s 145 tok/s 93 tok/s	~672 ms ~673 ms ~677 ms
#222	Llama 3.2 3B	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
#223	Phi-3-mini	3.8B	INT4 Q8 FP16	116 tok/s 77 tok/s 44 tok/s	~1.6s ~1.6s ~1.6s
#224	Gemma 3n E2B IT	6B (2B active)	INT4 Q8	90 tok/s 77 tok/s	~952 ms ~953 ms
#225	Gemma 2 2B	2B	INT4 Q8 FP16	167 tok/s 121 tok/s 74 tok/s	~887 ms ~889 ms ~894 ms
#226	Mistral-7B-Instruct-v0.1	7.3B	INT4 Q8	73 tok/s 45 tok/s	~3.0s ~3.0s
#227	Llama 3.2 1B	1B	INT4 Q8 FP16	228 tok/s 181 tok/s 123 tok/s	~460 ms ~461 ms ~464 ms
#228	Phi-3-medium	14B	INT4	42 tok/s	~5.6s
#230	OLMo 3 7B Instruct	7B	INT4 Q8	54 tok/s 36 tok/s	~2.9s ~3.1s
#231	Qwen2-7B	7B	INT4 Q8	75 tok/s 47 tok/s	~2.8s ~2.9s
#232	Phi-3-small	7B	INT4 Q8	75 tok/s 47 tok/s	~2.8s ~2.9s
#233	Mistral-7B-Instruct-v0.2	7.3B	INT4 Q8	73 tok/s 45 tok/s	~3.0s ~3.0s
#234	Gemma 3 1B	1B	INT4 Q8 FP16	228 tok/s 181 tok/s 123 tok/s	~460 ms ~461 ms ~464 ms
#236	Gemma 1 7B	7B	INT4 Q8	54 tok/s 36 tok/s	~2.9s ~3.1s
#237	Gemma 1 2B	2B	INT4 Q8 FP16	178 tok/s 126 tok/s 76 tok/s	~886 ms ~889 ms ~894 ms
#238	ChatGLM3-6B	6B	INT4 Q8	60 tok/s 43 tok/s	~2.5s ~2.5s
#239	ChatGLM2-6B	6B	INT4 Q8	60 tok/s 43 tok/s	~2.5s ~2.5s
#240	ChatGLM-6B	6B	INT4 Q8	60 tok/s 43 tok/s	~2.5s ~2.5s
Unranked Models (Sorted by Speed)
-	DINOv3 ViT-L/16	0.3B	INT4 Q8 FP16	315 tok/s 284 tok/s 233 tok/s	~136 ms ~136 ms ~137 ms
-	Qwen2-0.5B	0.5B	INT4 Q8 FP16	285 tok/s 245 tok/s 186 tok/s	~224 ms ~224 ms ~225 ms
-	Qwen2.5-0.5B	0.5B	INT4 Q8 FP16	285 tok/s 245 tok/s 186 tok/s	~224 ms ~224 ms ~225 ms
-	ERNIE-4.5-0.3B	0.3B	INT4 Q8 FP16	280 tok/s 257 tok/s 215 tok/s	~136 ms ~136 ms ~137 ms
-	ERNIE-4.5-0.3B-Base	0.3B	INT4 Q8 FP16	280 tok/s 257 tok/s 215 tok/s	~136 ms ~136 ms ~137 ms
-	Liquid D1-Omni 600M	0.6B	INT4 Q8 FP16	272 tok/s 229 tok/s 169 tok/s	~269 ms ~270 ms ~271 ms
-	Falcon-1B	1B	INT4 Q8 FP16	240 tok/s 187 tok/s 126 tok/s	~460 ms ~461 ms ~463 ms
-	Falcon3-1B	1B	INT4 Q8 FP16	228 tok/s 181 tok/s 123 tok/s	~460 ms ~461 ms ~464 ms
-	MiniCPM-V 4.7 1B	1B	INT4 Q8 FP16	228 tok/s 181 tok/s 123 tok/s	~460 ms ~461 ms ~464 ms
-	Qwen2-1.5B	1.5B	INT4 Q8 FP16	193 tok/s 145 tok/s 93 tok/s	~672 ms ~673 ms ~677 ms
-	Qwen2.5-1.5B	1.5B	INT4 Q8 FP16	193 tok/s 145 tok/s 93 tok/s	~672 ms ~673 ms ~677 ms
-	Spark X2.5 1.7B	1.7B	INT4 Q8 FP16	182 tok/s 134 tok/s 84 tok/s	~759 ms ~761 ms ~765 ms
-	Hy-MT2 1.8B	1.8B	INT4 Q8 FP16	177 tok/s 129 tok/s 81 tok/s	~799 ms ~801 ms ~806 ms
-	Falcon-3B	3B	INT4 Q8 FP16	142 tok/s 95 tok/s 55 tok/s	~1.3s ~1.3s ~1.3s
-	CroissantLLM Base	1.3B	INT4 Q8 FP16	140 tok/s 118 tok/s 85 tok/s	~588 ms ~589 ms ~592 ms
-	Phi-1	1.3B	INT4 Q8 FP16	140 tok/s 118 tok/s 85 tok/s	~588 ms ~589 ms ~592 ms
-	Phi-1.5	1.3B	INT4 Q8 FP16	140 tok/s 118 tok/s 85 tok/s	~588 ms ~589 ms ~592 ms
-	Mellum 2.1 12B A2.5B Thinking	12B (2.5B active)	INT4	137 tok/s	~1.2s
-	DeepSeek-R1 3B	3B	INT4 Q8 FP16	136 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	Falcon3-3B	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	LFM2-VL 3B	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	Liquid D1 3B	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	Liquid D1-VL 3B	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	Liquid LFM2-VL 3B D1	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	Qwen2.5-3B	3B	INT4 Q8 FP16	135 tok/s 92 tok/s 54 tok/s	~1.3s ~1.3s ~1.3s
-	FrogNano 4B	4B	INT4 Q8 FP16	113 tok/s 74 tok/s 42 tok/s	~1.7s ~1.7s ~1.7s
-	Spark X2.5 4B	4B	INT4 Q8 FP16	113 tok/s 74 tok/s 42 tok/s	~1.7s ~1.7s ~1.7s
-	EmbeddingGemma 2	2B	INT4 Q8 FP16	108 tok/s 88 tok/s 61 tok/s	~890 ms ~892 ms ~897 ms
-	Phi-2	2.7B	INT4 Q8 FP16	95 tok/s 75 tok/s 49 tok/s	~1.2s ~1.2s ~1.2s
-	MaLLaM-3B	3B	INT4 Q8 FP16	91 tok/s 70 tok/s 45 tok/s	~1.3s ~1.3s ~1.3s
-	SmolLM3 3B	3B	INT4 Q8 FP16	91 tok/s 70 tok/s 45 tok/s	~1.3s ~1.3s ~1.3s
-	Falcon-7B	7B	INT4 Q8	79 tok/s 48 tok/s	~2.8s ~2.9s
-	Falcon3-7B	7B	INT4 Q8	75 tok/s 47 tok/s	~2.8s ~2.9s
-	Hy-MT2 7B	7B	INT4 Q8	75 tok/s 47 tok/s	~2.8s ~2.9s
-	Mistral-7B-v0.1	7.3B	INT4 Q8	73 tok/s 45 tok/s	~3.0s ~3.0s
-	Youtu Parsing Omni	8B	INT4 Q8	70 tok/s 42 tok/s	~3.2s ~3.2s
-	LensVLM 9B	9B	INT4 Q8	62 tok/s 36 tok/s	~3.6s ~4.0s
-	MiMo V2.6 Distill Qwen 9B	9B	INT4 Q8	62 tok/s 36 tok/s	~3.6s ~4.0s
-	ChatGLM3-6B-32K	6B	INT4 Q8	60 tok/s 43 tok/s	~2.5s ~2.5s
-	Yi-6B	6B	INT4 Q8	60 tok/s 43 tok/s	~2.5s ~2.5s
-	Falcon3-10B	10B	INT4 Q8	57 tok/s 31 tok/s	~4.0s ~4.7s
-	Falcon2-11B	11B	INT4 Q8	54 tok/s 29 tok/s	~4.4s ~5.2s
-	Hunyuan Lite	7B	INT4 Q8	54 tok/s 36 tok/s	~2.9s ~3.1s
-	OLMo 3 7B Base	7B	INT4 Q8	54 tok/s 36 tok/s	~2.9s ~3.1s
-	SEA-LION-7B	7.1B	INT4 Q8	53 tok/s 36 tok/s	~2.9s ~3.1s
-	SEA-LION-7B-Instruct	7.1B	INT4 Q8	53 tok/s 36 tok/s	~2.9s ~3.1s
-	Sahabat-AI-Llama3-8B-Instruct	8B	INT4 Q8	49 tok/s 31 tok/s	~3.2s ~3.8s
-	Typhoon-2-8B	8B	INT4 Q8	49 tok/s 31 tok/s	~3.2s ~3.8s
-	GLM-4-9B	9B	INT4	45 tok/s	~3.6s
-	GLM-4-9B-Chat	9B	INT4	45 tok/s	~3.6s
-	GLM-4-9B-Chat-1M	9B	INT4	45 tok/s	~3.6s
-	Yi-9B	9B	INT4	45 tok/s	~3.6s
-	Sahabat-AI-Gemma2-9B	9.2B	INT4	44 tok/s	~3.7s
-	Sahabat-AI-Gemma2-9B-Instruct	9.2B	INT4	44 tok/s	~3.7s
-	GreenMind-14B-R1	14B	INT4	28 tok/s	~6.6s

Arc B580 (12GB)

VRAM Scaling by Context Length

Best Models to Run on Arc B580 (12GB)

Inference Capacity

Fine-Tuning Capacity

Workload Recommendations

Frequently Asked Questions