I keep asking myself the same question: can my GPU actually run this model?
Not in theory. In practice. I have an RTX 4090. I want to run Llama 3.3 70B. Do I need to buy another card? Should I settle for Q4 quantization? Will it even fit if I bump the context window to 32K?
Googling gets you Reddit threads with conflicting answers, or calculators that only tell you one side of the story. There’s a CLI called llmfit that’s actually pretty good — 29k stars on GitHub — but it runs on your machine, so it can only tell you what that machine can do. If you’re still deciding what to buy, it can’t help.
So I built something that can.
What GPU Fits Does
Three tools, all in the browser, no install needed.
VRAM Calculator
Pick a model, pick a quantization level (Q4_K_M, Q6_K, FP16, etc.), slide the context length. It breaks down the memory into three bars: weights, KV cache, and system overhead. The KV cache part is where it gets interesting — I had to handle both standard GQA models and MLA models like DeepSeek-R1 separately, because the latter compresses the cache so aggressively that the standard formula overestimates by orders of magnitude.
GPU Compatibility Checker (Bidirectional)
This is the one I use most. Two modes:
- Model → GPU: “I want to run Qwen3 32B at Q4 with 8K context — what cards can handle it?” Single card, multi-card, doesn’t matter. It sorts by whether it’s comfortable (≤80% VRAM), tight (≤100%), or needs a multi-GPU setup.
- GPU → Model: “I have a 4090 — what’s the biggest model I can comfortably run?” Or “I have two 3090s — what multiplies out to fit?”
Context length changes the answer. That’s the part most charts miss.
Local vs API Cost Calculator
The honest one. You plug in your expected monthly token usage, and it tells you whether buying hardware pays off compared to calling an API. It doesn’t assume “local always wins” — sometimes it doesn’t, and the tool says so. If the math says API is cheaper, it shows you the break-even point in months. No agenda, just numbers.
The Boring Tech Stack (Again)
Same as NZ Pathway: Astro 4, Tailwind, static HTML. I cloned the skeleton from the immigration project and stripped out all the NZ content. The i18n pipeline, Cloudflare Pages deployment, and SEO setup were already battle-tested, so this one went faster.
The whole site is bilingual (English + Chinese), because a lot of the people building with local LLMs are in both ecosystems.
One thing that’s not boring: the site auto-generates 138 combination pages. Like “can an RTX 4090 run Llama 3.1 70B?” Each page has real computed numbers, not copy-pasted text. That’s 138 × 2 languages = 276 pages, all static HTML, built at deploy time from a 12×12 model×GPU matrix. The idea is to catch long-tail search queries that nobody writes content for.
Where We Are
Everything’s built. 313 pages total. 20/20 test vectors pass (I wrote unit tests for the VRAM math because getting that wrong would be worse than not having a calculator at all). The data layer covers 12 models, 12 GPUs, 7 quantization presets, and pricing from 7 API providers.
Domain’s bought, Cloudflare Pages is wired up, and the site’s live.
Same rules as the last project. Tools first, trust second, ads whenever. And I’m still not pretending to be anything I’m not — there’s a disclaimer on every page saying these are estimates, not guarantees, and you should check the numbers yourself.
It’s live at gpufits.com. If you know the pain of calculating whether your GPU can fit a 70B model, you might find it useful.