About VRAMCalculator.net

VRAMCalculator.net helps people choose hardware for running AI models on their own computers. It answers one question as precisely as possible: will this model run on this GPU or Mac, and how well? The site is run by an independent publisher. It is written for people running open models on consumer GPUs, Macs, unified-memory mini PCs and rented cloud servers.

Why this site exists

Most VRAM advice online is a rule of thumb such as "about 2 GB per billion parameters", repeated without context. Those rules break for exactly the models people care about today. Mixture-of-experts models need memory for every expert but only read a few per token. Gemma uses sliding-window attention, so long contexts are cheap. Qwen 3.5 and later keep a KV cache on only a quarter of their layers. DeepSeek and Kimi compress the cache into a small latent vector. A flat percentage gets all of these wrong, and the mistake usually shows up after someone has already bought a graphics card.

We built calculators that use each model's real architecture instead. Every figure on the site comes from the same formulas, so a number on a model page, a GPU page, a guide and the calculator always agree.

Where the numbers come from

• Model architecture: layer counts, attention heads, key/value heads, head dimensions, vocabulary size, sliding windows, expert counts and attention types are copied from each model's official config.json on Hugging Face. Each model page links to its source.
• GPU specifications: memory size and bandwidth come from manufacturer specification sheets. For Macs and unified-memory systems we use the default share of memory the GPU can use.
• Formulas: weights, KV cache, runtime overhead, fit thresholds and speed estimates are documented in full on the methodology page.
• Calibration: we compare our estimates against memory and speed figures reported by llama.cpp, Ollama, LM Studio and vLLM users, and adjust the formulas when they drift. The speed model was recalibrated in October 2026 after mixture-of-experts estimates ran high.

Keeping it accurate

New models are usually added within days of release. Every data or formula change is logged with a date on the changelog, and the complete data set is public as JSON at /data/models.json and /data/gpus.json, so anyone can check our work. Where a figure is an estimate rather than a published value, such as the native file size of a model whose checkpoint size is not stated, the page says so.

Corrections

If your runtime reports noticeably different memory use or speed, we want to know. Email hello@vramcalculator.net with the model, quantization, context length, runtime and the number you saw. Confirmed errors are fixed and recorded in the changelog.

How the site is funded

The calculators are free and need no account. The site may show display advertising, which pays for hosting and the time spent maintaining the data. Advertisers have no influence on our figures, comparisons or recommendations, and we do not accept payment for placement in comparisons or guides. If we ever add affiliate links, they will be clearly labelled.

Coverage today

58 language models, 15 image and video models and 69 GPUs and unified-memory systems, plus 18 guides. Missing something? Ask us to add it.