Best models for your card
How the math works
Three things compete for your card's memory:
- Model weights — the model's parameter count times the bits each parameter is stored at. Compression ("quantization") shrinks this: the same model at Q4 needs about 70% less memory than at full precision, with only a small quality cost.
- Context memory — every token of your conversation is cached so the model doesn't re-read it. This grows linearly with the context window and differs per model architecture; long-context work on big models can eat several gigabytes on its own.
- Overhead — the runtime's working buffers (~0.6 GB), and if the same card drives your monitor, the desktop's slice too (~0.8 GB).
The verdicts: comfortable means at least 10% headroom. Tight means it fits but leave the browser tabs closed. Shorten the context means the weights fit but your chosen context window doesn't. All figures are planning estimates — real usage varies a little by runtime and driver.
Mixture-of-experts models (like Mixtral or gpt-oss) are fast per token, but all experts must sit in memory — we count the full size, as your card must too.
Pricing
Fit checker
- Instant fit verdicts for 18 popular open models
- Full memory breakdown (weights, context, overhead)
- Best-models shortlist for your card
- No account, no tracking, nothing uploaded
Pro planning report
- Best-value GPU upgrade paths for your budget
- Exact model + compression shortlists for the card you choose
- Memory math for fine-tuning, not just running
- Multi-GPU and unified-memory (Mac) planning
Written for your exact card and emailed within 24 hours. Not happy with it? Reply to your receipt and you get your money back.
FAQ
My card isn't listed.
Pick "Custom / enter memory manually" at the bottom of the card list and type your card's memory in gigabytes. Everything else works the same.
The verdict says "no fit" — what are my options?
In order of preference: choose a stronger compression level (Q4 is the sweet spot), pick a smaller model from your shortlist, or shorten the context window. Most runtimes can also spill layers to normal RAM — it works, but expect a large slowdown.
How accurate are these numbers?
Within a few percent for the weights (they match published file sizes) and a close estimate for context memory based on each model's architecture. Runtimes differ slightly in overhead, which is why the checker keeps a safety margin before calling anything "comfortable".
Is anything sent to a server?
No. The whole checker is a static page — every calculation happens in your browser. There's no account, no cookie banner because there are no cookies, and nothing to upload.