Hi 👋 I'm the person behind ai-muninn 🐦⬛
Local-AI infra, built and broken at home. I quantize LLMs to FP8 / NVFP4, run them on Blackwell (RTX 5090 / DGX Spark GB10), wire up AI agents, and write down everything I trip over — so you don't have to.
What I'm into
- 🔧 LLM quantization (FP8 / NVFP4), vLLM & llama.cpp serving, speculative decoding, multimodal
- 📏 Measure-first benchmarking — I publish the numbers, including the ones that don't pay off
- 🤖 Local AI agents & the tooling around them
🤗 Models — quantized builds on Hugging Face, 28k+ downloads and counting:
- Gemma-4-E4B-it-NVFP4A16 — 11.5k
- gemma-4-12B-it-FP8-dynamic — 9.3k
- Huihui-Qwen3.6-35B-A3B-abliterated-FP8-DYNAMIC — 2.7k
📝 Writeups → ai-muninn.com — the traps, the benchmarks, the stuff the vendor blog won't tell you.