Hey PH 👋 I built QuantaMind because I kept getting burned by local models that looked great on paper and fell apart the moment I put them in an agent loop. The pattern was always the same: a model scores well on some public leaderboard, I load it locally, and then it quietly breaks, a malformed tool call at step 7, or worse, it says "task complete" when nothing got done. Public benchmarks test the model in the cloud. They don't test your GPU, your VRAM, or your quantization, and that combination is exactly what breaks. The question I heard everywhere (and asked myself constantly) was: "I have X GB of unified memory / vRAM, which model actually works?" The honest answer today is a month of trial and error. QuantaMind runs the real agent loop on your own hardware, fully offline, and gives you a Ready / Conditional / Not Ready verdict per model and quant — with the reasons named, never a bare score. It works with Ollama, llama.cpp, and MLX. A few things I care about that shaped it: - It verifies the actual tool calls happened, not the model's claim of "done" — a trivial do-nothing agent scores ~0, not the inflated numbers you see on some benchmarks. - It never fabricates a metric. If something can't be measured on your backend, it says N/A. - It's fully open source and runs entirely on your machine. no prompts or data ever leave. It's early and I'd genuinely love feedback from people running local models day to day. If you've hit the "fits in memory but still fails" wall, I'd love to hear what model/hardware combo burned you, and contributions are very welcome. One thing I'm curious about: what's the smallest quant you've actually trusted in a real agentic workflow?

QuantaMind

Can your local model actually run an agent? Test it - OSS.

Can your local model actually run an agent? Test it - OSS.