Smeltcore
Which AI models actually run on your GPU
Start new thread

Smeltcore - Which AI models actually run on your GPU

by•
Honest benchmarks for running open-weights AI models on your own GPU. 84 models × 27 consumer GPUs — from RTX 3060 to Apple Silicon — each pair gets an honest verdict: runs, tight fit, or won't fit, with real speed and VRAM numbers. Plus 800+ step-by-step install recipes (llama.cpp, Ollama, ComfyUI, MLX) and a GPU Advisor. No invented numbers, no signup, free under CC BY-SA.

Add a comment

Replies

Best
Maker
📌
Hi Product Hunt 👋 Every time I wanted to run a new open model locally I hit the same wall: "will this actually fit on my GPU?" The real numbers were out there — but scattered across Reddit threads, blog posts and GitHub issues, half of them untested. Mostly you'd just get vendor marketing, guesses, or "just buy a 4090." So I built Smeltcore: those numbers pulled into one place, each linked to its source and sanity-checked. Pick your GPU (RTX 3060 → 4090, or Apple Silicon) and see every model it can run — with real speed and peak-VRAM numbers and a plain verdict: runs, tight fit, or won't fit. 84 models, 27 GPUs, 800+ step-by-step install recipes for llama.cpp, Ollama, ComfyUI and MLX. Free, no signup, CC BY-SA. No invented numbers. Every benchmark links to its source, "doesn't fit" is a first-class answer, and community submissions go through a review gate. It started as a spreadsheet for myself and turned into this. Would love your feedback — especially which model × GPU pairs you're missing. And if you've benchmarked something, there's a /contribute form 🙏

finally a benchmark site that doesn't oversell. checked my 4070 against a few models and the VRAM numbers actually matched what ollama shows locally. the install recipes are a nice touch too.

Maker

 Thank you for your feedback!

Finally a benchmark site that doesn't oversell performance. Checked the RTX 3060 vs 7B models section and the VRAM numbers matched what I actually get running llama.cpp locally. That kind of honesty is refreshing.

the "tight fit" verdict is genuinely useful, like knowing exactly when you'll be sweating at 99% VRAM is way more honest than those blanket "it runs" claims. and no signup for the full thing under CC BY-SA is a really solid move.

A repo of actual install recipes and real VRAM numbers is genuinely useful, way better than guessing from random Reddit threads. One idea if you're open to it: add a simple toggle or filter to compare models head-to-head on the same GPU, like side-by-side tokens/sec for Llama 3 8B vs Qwen 2.5 7B on a 4070. Would make picking between similarly sized models a lot less of a guessing game.

Love the idea behind Smeltcore. Big congratulations on shipping and launching! 🚀