Memory planning / LLM VRAM calculator
Make room
for your model.
See how weights, context and runtime add up. Compare the estimated footprint with the memory you can actually use.
Estimated device memory
- Model weightsQ4_K_M (4-bit) · 4.85 effective bits / weight
- 4.8 GB
- KV cache8k tokens · generic FP16 cache estimate
- 1.0 GB
- Runtime allowanceFixed allowance in this planning model
- 1.5 GB
Within the modeled memory
The estimate is below your usable memory. This does not guarantee that a model will load or perform well.
- System RAM to load estimate
- 8 GB
- Model file size estimate
- 5.1 GB
Arithmetic over stated assumptions, not a benchmark. Architecture, KV heads, cache precision, runtime and offloading can change the actual footprint.
Memory is a constraint.
The model is an approximation.
Memory capacity affects whether a model and its runtime allocations fit on a device. Speed also depends on memory bandwidth, compute, architecture and the inference runtime. This calculator estimates capacity; it does not predict token speed.
- Weights
- Parameter count × effective bits per weight ÷ 8. Format overhead is included in the effective bit values; the selected format is not a quality guarantee.
- KV cache
- A generic FP16 cache estimate using parameter-count bands and context length. Actual KV-head counts and cache precision vary by model and runtime.
- Runtime
- A fixed 1.5 GB allowance. Real allocation depends on the framework, device and workload. Leave additional room when planning deployment.
The input ranges are calculator bounds. They do not establish supported model context or hardware configurations.
Start from a known size.
Open a model page for parameter counts, references and precomputed scenarios.