Model guides · updated 2026-09-10

What it costs to run each model

Per-model guides connecting VRAM requirements to this week's cheapest graded GPU rentals. Pick your model; get GPU counts and live prices at every precision.

Llama 3.1 8B

≈5 GB at 4-bit · from $0.37/hr

Phi-4 (14B)

≈8 GB at 4-bit · from $0.37/hr

Mistral Small 3 (24B)

≈14 GB at 4-bit · from $0.37/hr

Gemma 2 27B

≈16 GB at 4-bit · from $0.37/hr

Qwen2.5 32B

≈19 GB at 4-bit · from $0.37/hr

Llama 3.3 70B

≈42 GB at 4-bit · from $0.75/hr

Qwen2.5 72B

≈43 GB at 4-bit · from $0.75/hr

Llama 4 Scout (109B MoE)

≈65 GB at 4-bit · from $0.91/hr

gpt-oss-120b (117B MoE)

≈70 GB at 4-bit · from $0.91/hr

Mistral Large 2 (123B)

≈74 GB at 4-bit · from $0.91/hr

Mixtral 8x22B (141B MoE)

≈85 GB at 4-bit · from $1.20/hr

Qwen3 235B (MoE)

≈141 GB at 4-bit · from $1.81/hr

Llama 4 Maverick (400B MoE)

≈240 GB at 4-bit · from $2.72/hr

Llama 3.1 405B

≈243 GB at 4-bit · from $3.21/hr

DeepSeek-R1 (671B MoE)

≈403 GB at 4-bit · from $5.21/hr

Kimi K2 (1T MoE)

≈616 GB at 4-bit · from $7.26/hr

Any other size: the interactive fit finder covers arbitrary models and precisions. Thinking in tokens instead of GPUs? What a million tokens actually costs.