NeuronDump: open-source tool to inspect and shrink local LLMs

A developer released NeuronDump, an MIT-licensed Rust binary with a CLI and local web UI that parses GGUF files directly. It visualizes layers, attention heads, vocabulary and chat templates, and lets users cut parts of a model to produce a smaller GGUF that still loads in llama.cpp.

• Reads GGUF files directly with nothing hard-coded per model
• Sweep mode ranks layers by damage and writes a smaller GGUF plus an Ollama Modelfile
• Trace feature compares the cut model against the original via KL divergence and perplexity
• Supports llama, qwen2/3, gemma, phi3 and bert families; inference is CPU-only