NeuronDump: inspect, edit and slim any GGUF, with a 3D model graph (CLI + web UI)

[aymen94]

I built NeuronDump, a single binary with a CLI and a local web UI that reads GGUF directly: https://github.com/aymen94/NeuronDump

• Inspect: every layer, head, tensor and quant type, plus the tokenizer and chat template
• 3D graph: open the model from layers down to single heads, neurons and experts
• Trace: follow a token through every layer (logit lens, attention per head, which heads decided it)
• Edit: zero, scale or skip parts, steer, ban tokens, and compare live against the original
• Slim: cut parts, see the impact, and write a new GGUF that still loads in llama.cpp

Found so far: DeepSeek-R1 8B has the same neurons in the same places as Qwen3 8B, and DeepSeek's template emits tool-call markers that aren't in its vocab (each one becomes 11 tokens).

CPU-only for now. If it writes a GGUF that llama.cpp rejects, please tell me.