The Broke Vibe Coder

Clover-35B-A3B Kapitsa 4-bit pack and DFlash2 drafter quick benchmarks on a m5 pro Mac mini 64GB.

A quantization explanation nobody asked for: Think of LLM quantization like audio compression. The more you compress a song, the less detail you preserve. But even when the quality drops, you can still recognize the music. Here’s my completely unscientific audio analogy: 📻 Q1 — AM radio on an ’80s boombox 📻 Q2 — FM baseball radio 📼 Q3 — Cassette tape 💽 Q4 — Sony MiniDisc 🎧 Q5 — Modern high-quality MP3 🔊 Q6 — FLAC on a home theater system 🎼 Q8 — FLAC on a $100,000 stereo 🎤 FP16 — Front row at a live concert 🎹 FP32 — In the studio with the band The song stays the same. What changes is how clearly you can hear every detail. Your LLM works in a similar way. It still knows the melody. It can follow the rhythm, recognize the instruments, and play the song. But somewhere along the way, it might lose the subtle shimmer of a cymbal, the harmonies of a backup singer, or the faint sound of a guitar buried in the mix. And that’s the catch with quantization. The model doesn’t necessarily forget the song. It loses some of the precision needed to reproduce every detail. That’s why heavily quantized models can still work surprisingly well, but occasionally struggle with the finer distinctions. Sometimes those missing details don’t matter. Sometimes they’re the difference between a perfect performance and a wrong note. Not a literal technical equivalence, obviously. Just a way to understand why fewer bits can still produce something remarkably close to the original.