Model Hub: VRAM Requirements & GPU Fit

Each model page includes architecture specs, five-quant VRAM breakdowns, a verdict matrix across 12 GPUs (comfortable / tight / needs multi-GPU / infeasible) and theoretical speed estimates. All numbers are computed with our formulas: weights = params × bpw ÷ 8, KV cache at 8K context in fp16, plus 1.5GB runtime overhead.

Model	Parameters	Architecture	Q4_K_M @8K VRAM
Llama 3.2 3B	3.21B	Dense	4.4 GB
Llama 3.1 8B	8.03B	Dense	7.5 GB
Qwen3 8B	8.2B	Dense	7.7 GB
Phi-4 14B	14.7B	Dense	12.2 GB
gpt-oss-20b	20.9B / 3.6B	MoE	14.7 GB
Mistral Small 3.2 24B	24B	Dense	17.5 GB
Gemma 3 27B	27.4B	Dense	22.4 GB
Qwen3.8 27B	27.8B	Dense	20.7 GB
Muse Glimmer 30B	29.6B	Dense	20.1 GB
Qwen3 30B-A3B	30.5B / 3.3B	MoE	21.0 GB
Qwen3 32B	32.8B	Dense	23.7 GB
Llama 3.3 70B	70.6B	Dense	47.4 GB
gpt-oss-120b	116.8B / 5.1B	MoE	73.6 GB
DeepSeek-R1 671B	671B / 37B	MoE · MLA	413.1 GB

Data verified 2026-09-01