Can my PC run this model?
Tell us your rig. We'll tell you what runs.
No VRAM math, no guessing. Enter your machine once and instantly see every popular model scored — what runs great, what runs with a tweak, and what your hardware simply can't hold. We even model RAM offload and Apple unified memory, which almost nobody does.
—
Memory is the model's weight footprint plus overhead at the best-fitting quant. Speed is the bandwidth-bound decode ceiling on active parameters — a clean upper bound, lowered honestly when layers spill to RAM. Context is the max we can fit in leftover memory.
Why this one is different
Most tools stop at “does it fit.”
OS-aware VRAM
Windows reserves memory the driver won't give back; Linux gives you the most; macOS shares one pool. We compute usable VRAM per OS, not the sticker number.
RAM offload, modelled
When a model doesn't fit VRAM, real tools spill layers to system RAM — slower, but it runs. We tell you it'll run and the honest speed hit, instead of a flat “no.”
Apple unified memory
On Apple Silicon, RAM is VRAM. Pick macOS and a 64 GB Mac runs models a 24 GB card can't. Almost no calculator gets this right.
Know what runs?
Now cost it out.
See what the card that runs your model actually costs to own — and whether renting wins.
Evidence & method
How this calculation works
Checks each model's estimated memory footprint against your GPU's usable VRAM and system RAM, flagging fit, tightness, RAM offload and a recommended quantization.
Data sources
- NVIDIA, AMD & Intel GPU documentation
- Hugging Face model cards
- Official model papers
- Bitpute Methodology
Assumptions
- Weights = parameters × bits-per-weight ÷ 8
- Runtime overhead = 0.75 GB + 5% of weights (CUDA context, allocator, buffers)
- Recommended VRAM = total × 1.10 (10% headroom)
- Usable VRAM = card capacity × 0.95
- KV cache counted only when the model architecture is known; when applied, batch size 1, FP16 KV, context as entered
- Single GPU, inference workload
Limitations
- Actual VRAM varies by framework (PyTorch, llama.cpp, vLLM, TGI), driver/CUDA version and OS
- Quantization implementations differ; real bits-per-weight can vary from the nominal value
- When architecture is unknown the KV cache is omitted, so long-context use will exceed the estimate
- Estimates are for planning and comparison, not a guarantee
Data status: Hardware specs — vendor documentation. Model metadata — community model cards, not independently verified.
Related
Why does this estimate differ from other calculators?
- Different footprint formulas and rounding
- Different KV-cache assumptions (batch, context, precision)
- Framework and inference-engine differences
- Reserved and fragmented VRAM
- Driver and CUDA overhead
- Precision and quantization choices