Updated 8 Oct 2026 1 new this month

Can your hardware run it? The local LLM VRAM calculator.

Find the open-weight models your computer can actually run — and which one to pick for coding, reasoning, vision or chat. Choose your GPU or Apple Silicon chip: we size the weights at your quantisation, add the KV cache at your context length, estimate tokens per second, and recommend a model per job.

Common rigs RTX 4090 · 24 GBRTX 3060 12 GBRX 7900 XTX · 24 GBM3 Max · 64 GB
Paste any Hugging Face model

Any org/model id or hf.co link — not just the 80 we have verified. We read its config.json, work out the parameters, layers, KV heads, context window and experts, and size it against your GPU. Ollama tags and model names work too.

Try: Qwen3.8 27B · Gemma 4 31B · gpt-oss 20B · an uncatalogued one
Reverse question: what hardware do I need for Qwen3.8 27B?

New this month

1 models released in the 31 days before 8 Oct 2026 · all models, newest first
DeepSeek V4.1 Flash 552B-A16BNew

Twice V4 Flash's backbone (552B, 16B active per generated token) plus 196B of Engram lookup tables (189 GiB at FP8) that DwarfStar streams from SSD, so they are not counted here. A 512 GB Mac model, and as of October 2026 no mainline llama.cpp or Ollama build: DwarfStar on a Mac, vLLM across four GPUs. DeepSeek puts the cache at 890 bytes/token; modelled conservatively as 4 latent layers plus a 128-token window.

DeepSeek · 10 Sep 2026 · MIT

91

GPUs & chips profiled

80

Models sized

1

New in the last month

5

Runtimes covered

computed

Not benchmarked

Start here

Pick the device, then the job

VRAM is the binding constraint, and it is the one number you already know. Choose your card and you get a recommendation per use case — best overall, coding, reasoning, vision, agents, fastest — plus every model in the catalogue sorted into runs-great, runs-tight and won't-fit.

Browse all 91 devices →

The trap

The weights are only half of it

A 32B model at Q4_K_M is 18.4 GB of weights, which looks fine on a 24 GB card — until you ask for 128K context and the KV cache alone wants 32 GB. Every table here prices the cache at the context you actually chose.

How the sizing works →

Then

One command, not a weekend

Every model page ends in the exact command for Ollama, llama.cpp, LM Studio or MLX, with the context flag already set to the size we said would fit.

Compare the five runtimes →

Start with a question

Best local LLM for my VRAM

One pick per job for every memory class, 8 GB, 12 GB, 16 GB, 24 GB, 32 GB, 48 GB and the 64 and 128 GB Macs — plus the coding shortlist, ranked.

Which GPU should I buy?

Every card and Mac ranked by what it can run and how fast, with the smallest card each model needs. Then how to run an LLM locally, from hardware to the command.

How much VRAM does it need?

The VRAM calculator and KV cache calculator for any model and context, what Q4_K_M means, and which Ollama models fit your card.

Not sure local is worth it yet? The honest comparison is cost per unit of work: a GPU is one payment, an API is a rate per million tokens. The AI article price calculator costs a finished article across 40 hosted Claude, GPT-5, Gemini and Grok models — hold that against the card you were about to buy.