Will this model fit in my GPU memory?
GPU memory estimation using the full formula — weights, KV cache (GQA-aware), activations, and CUDA overhead.
Advanced tool
Answers: Whether a specific model configuration is estimated to fit in your GPU's memory.
Does not answer: Whether the model will run well, feel fast, or suit your workload.
This estimates GPU memory only. It does not prove the model will run well, feel fast, or suit your workload.
This is a GPU memory estimate. It is not a claim that the model will run, run well, or feel fast.
We earn commissions when you shop through the links below at no additional cost to you.
Need more VRAM?Deploy on a cloud GPU. Pricing and availability change; check the provider for current details.
Get the RigFact welcome message.
One welcome email, with where to start — no spam.
Did this estimate match your real experience?
Thanks! This helps us improve. 🙏
Why isn't my model listed?
What this calculator covers
Only models whose architecture parameters are source-verified.The calculator only estimates models whose architecture parameters are source-verified. A current catalog model appears here once its architecture is verified — it is not missing because it is old, and its absence is not a fit or quality verdict.
17 current 2026 catalog models are not estimated here yet — their architecture parameters are not source-verified, so no estimate is invented for them: Qwen3.8 27B · Qwen3.6 27B · Qwen3.6 35B A3B · Muse Glimmer 30B · Qwen3-Coder 30B A3B · Gemma 4 26B MoE · gpt-oss 20B · Mistral Small 3.2 24B · DeepSeek-R1-Distill-Qwen 32B · GLM-4.7-Flash · GLM-5.3 · Kimi K3 · DeepSeek V4-Pro · Qwen3.5 397B · Llama 4 Scout · gpt-oss 120B · Mistral Medium 3.5. They stay discoverable under Models relevant to local AI in 2026 and in the model catalog. Absence from the calculator is not a currency, fit or quality verdict.
Tested RigFact Recipes are a third, separate thing: executed configurations with captured evidence. A calculator estimate is not a Recipe, and a catalog entry is neither.
How this estimate works — technical assumptions
The formula behind the estimate
params × bits_per_param ÷ 8 ÷ 1.07
Q4_K_M ≈ 4.5 bits → 70B × 4.5/8 ≈ 39.4 GiB
2 × n_layers × n_kv_heads × d_head × context × 2 bytes ÷ 1.07
Llama 3: 8 KV heads (not 64) = 8× less KV cache
~1 GB inference / ~3 GB activations (fine-tuning)
Forward pass activations only — full fine-tuning needs significantly more (gradients + optimizer states).
CUDA context + engine overhead
llama.cpp ≈ 0.8 GB, vLLM ≈ 3.5 GB