How much VRAM does an LLM actually need?

Three things add up: the model's weights — we use measured file sizes from published builds, not parameter math; the KV cache, which grows with context length and follows the model's attention architecture; and a flat ~0.5 GiB of runtime overhead. Every component on a model page carries an evidence tag: measured, reported, or provisional.

For most people, a 24GB card — RTX 3090 or 4090 — is the sweet spot: the 27B-class dense models and 30B-class MoE models in the table above run at Q4_K_M with usable context where a Q4 build exists (the two Bonsai models are low-bit only). Four frontier models at the bottom exceed any consumer card; their pages price out the cloud and appliance exits instead. Llama 4 Scout is the near miss: its smallest measured build fits an RTX 5090 (32GB), tight.

Four tracked models exceed every consumer card at every measured build — Kimi K3, GLM-5.3-Flash, and both DeepSeek-V4-Flash versions. Each of their pages prices the realistic exits instead of GPU tiers: cloud rental (DeepSeek-V4-Flash's smallest door is a single 192GB B200), a 512GB unified-memory appliance, a distillation ladder, private deployment, or the API. The near miss is Llama 4 Scout: its smallest measured build fits an RTX 5090 (32GB) tight, so it only needs the cloud if you don't have a 32GB card.

Yes, for small models. The phone pages track 3–8B models at Q4_K_M with 4K context: a 12GB-class phone runs Gemma 4 E4B (tight), an 8GB-class phone runs Gemma 3n E4B razor-tight, and 6GB-class phones run Qwen3-4B, Phi-4-mini, or Llama 3.2 3B. Phone math counts usable RAM at 60% of the sticker number — each phone page carries the per-build table and measured file sizes.

Start from a published build: take the measured file size for your quantization tier, add KV cache for your context length, and add runtime overhead. On any model page here, that breakdown is itemized per component with pinned revisions and an evidence tag on each number.

A GPU whose VRAM fits the model at your quantization tier (or a unified-memory machine), a runtime such as llama.cpp or Ollama, and a published build of the model. Each model page lists the exact builds and revisions we measured, plus the cheapest hardware each one fits.

About 6 GB of RAM is the entry door and 12 GB is comfortable. At 4K context the tracked phone builds need roughly 2.8–5.2 GiB of free RAM at Q4_K_M, and phone math counts usable RAM at 60% of the sticker number — the OS and your background apps keep the rest. Below the line there is no swap: the app dies instead of slowing down. The phone table above lists which model class each RAM tier runs.

Some, and at 4-bit less than most people expect — that trade is exactly why Q4_K_M is the community default quality/size point: roughly a quarter of the 16-bit file size for a small quality step. Below about 3 bits the loss grows fast, and 1-bit builds are a last resort for fitting, not a default. Each model page's quant table marks the recommended point.

Yes — Apple Silicon Macs run local LLMs from unified memory, which acts like VRAM: the same fit math applies (measured file size + KV cache + overhead) against the machine's total unified memory. A 24 GB Mac behaves roughly like a 24 GB card, and 32–128 GB machines open doors consumer GPUs cannot. The largest tracked models exceed every consumer card — even the biggest only fit a maxed-out 512 GB Mac Studio at 1-bit (wired limit raised) — so their pages price the cloud and appliance exits first. The Mac page measures every memory tier's pick and ceiling across every tracked build.