Use this LLM VRAM calculator to turn a local-AI workload into an explainable hardware shortlist. It estimates a planning range, not a guaranteed fit or speed. Your exact model build, runtime and settings remain authoritative.
Need a quick model-class starting point? Read what size LLM can I run on my GPU?, then return here to calculate the exact precision, context and concurrency.
Not sure what to calculate yet? Start with the Can my PC run a local LLM? four-check hardware and workload guide.
Estimating a 13B model? Use the 13B vs 14B requirements guide to compare raw weights with architecture-specific cache geometry.
Estimating a 27B model? Use the 27B vs 32B requirements guide to account for the exact artifact, hybrid cache and multimodal allocations.
Planning Qwen3.8-27B? Stage the quick scenario below for a generic 27B, 4-bit, 8K, one-job estimate, then compare it with the exact 16.464GB community Q4 file or the shared-memory Mac guide. The preset does not calculate, save or transmit anything.
No scenario staged. Your platform and budget stay under your control.
Choose your workload and constraints, then build a shortlist.
LLM VRAM calculator reference
Raw model weights are approximately parameters × bits per weight ÷ eight. The table is arithmetic, not a fit guarantee. Add memory for the KV cache, runtime buffers, multimodal components, the operating system and concurrent requests.
| Model size | 4-bit weights | 8-bit weights | 16-bit weights |
|---|---|---|---|
| 7B | 3.5GB | 7GB | 14GB |
| 13B | 6.5GB | 13GB | 26GB |
| 32B | 16GB | 32GB | 64GB |
| 70B | 35GB | 70GB | 140GB |
A downloaded file can fit while the running workload does not. Read how context length changes KV-cache memory, and use the calculator above for a deliberately conservative planning range.
How the estimate works
The worksheet starts with model parameters × bits per weight ÷ eight. It then adds a visible planning buffer that grows with context, concurrency and workload type, plus 3GB of fixed runtime headroom. The result is rounded up to the next 4GB capacity step. This deliberately favors explainability over false precision.
The estimate does not model every architecture’s KV cache, multimodal projector, speculative-decoding setup, optimizer state, framework allocation strategy or operating-system reservation. Fine-tuning and training needs can exceed the estimate substantially. Treat the result as a shortlist gate, not a compatibility guarantee.
What a recommendation means
- A candidate must clear the estimated memory range, the selected platform constraint and the planning-price floor.
- Planning-price floors are broad screening thresholds, not current offers or promised purchase prices.
- The tool does not rank sellers, predict benchmark performance or assume unknown shipping is free.
- Multi-GPU capacity is excluded from automatic recommendations because separate VRAM pools are not universally additive.
For sourced hardware categories and compatibility caveats, use the GPU VRAM chart. For the marketplace evidence policy, read how we evaluate hardware deals.
Need a second look? Ask with your goal, current hardware and blocker. If you test the recommendation, share the result, or follow the weekly update by RSS.
Take it back to the community
Help improve the next result. New to this? Ask with your goal, current hardware and blocker. Experienced readers can share a reproducible build or useful finding, or flag a source or record correction.
Find the weekly workshop, then return to your original thread with the outcome. A helpful contribution earns thanks, not an automatic evidence label.
Planning a 70B-class model? Read the 70B LLM VRAM requirements guide for exact artifact-size and KV-cache examples across 24GB to 96GB memory tiers.
Not sure which “4-bit” or “8-bit” file to choose? Read the Q4_K_M vs Q5_K_M vs Q6_K vs Q8_0 guide; the label alone does not equal total file size, VRAM use, quality or speed.
Planning to exceed GPU VRAM? Read the CPU-offload guide to separate model fit from system-RAM pressure, device transfers, KV-cache placement and usable speed.
Considering Apple silicon? Read the Mac Studio unified-memory guide to translate the planning range into a current M5 Max or M5 Ultra configuration with OS headroom.
\nPlanning a 32B or 33B model? Use the 32B VRAM requirements guide for official Qwen artifact sizes, KV-cache arithmetic and 16GB-through-64GB tier tradeoffs.
\nPlanning a 14B or 15B model? Use the 14B VRAM requirements guide for official Qwen artifact sizes, KV-cache arithmetic and 8GB-through-24GB tier tradeoffs.
How much VRAM do you need for a 7B LLM?
Compare 4GB through 16GB tiers using official Qwen2.5-7B GGUF sizes, explicit KV-cache arithmetic and complete-workload headroom.
Choosing between 8GB, 12GB and 16GB?
Compare mainstream memory tiers using official 7B and 14B artifact sizes, context headroom and exact GPU-variant checks.
RTX 3060 12GB vs 4060 Ti 16GB vs 5060 Ti 16GB
Compare mainstream NVIDIA paths by exact capacity, architecture, CUDA generation, power and verified software fit—not gaming charts or peak metrics.
Sizing an 8B model?
Compare the calculator range with official Qwen3-8B GGUF file sizes and explicit KV-cache arithmetic.
Sizing a 4B model?
Compare the calculator range with official Qwen3-4B artifact sizes and explicit KV-cache arithmetic.
Llama 3.2 reference: compare the calculator’s planning range with Meta’s official 1B, 3B, 11B Vision and 90B Vision checkpoint sizes and exact cache geometry.