LLM VRAM Calculator
How much GPU memory do you need to run or fine-tune a model? Pick the size, precision, and what you're doing — get a VRAM estimate and a GPU that fits. Everything runs in your browser.
Estimate only — actual usage varies with batch size, sequence length, framework, and kernels. For models that use grouped-query or multi-query attention (most 70B+ open models), the KV-cache figure here is a deliberate over-estimate, so real inference memory is often lower.
Related guides
- LoRA & QLoRA fine-tuning guide — how the modes above actually differ
- LLM inference engines compared — once it fits, serve it efficiently
- Self-hosting LLMs vs cloud APIs — is buying the GPU worth it?
Related reading
Fine-Tune LLMs with LoRA and QLoRA: 2026 Guide
Learn to fine-tune LLMs with LoRA and QLoRA in 2026. VRAM requirements, dataset prep, Unsloth/Axolotl setup, hyperparameters, and evaluation.
Read →SpecKV: Adaptive Speculative Decoding with Dynamic Gamma
SpecKV (arXiv:2605.02888) shows fixed γ=4 costs 56% throughput. Adaptive gamma, KV cache compression effects, and vLLM production tuning guide.
Read →Self-Hosting LLMs vs Cloud APIs: Cost, Speed, Privacy 2026
Self-hosting LLMs with Ollama, vLLM, and llama.cpp vs cloud APIs: cost-per-token modeling, hardware needs, latency, and when each approach wins.
Read →