NeuronDump

Look inside any local LLM, find the parts it doesn't need, and make it smaller.

NeuronDump reads GGUF model files (the format used by Ollama and llama.cpp), shows what's inside them, and can write a smaller version that still runs.

What it does

Command	You get
list	all models on your computer
inspect	architecture, layers, heads, sizes, warnings
vocab	the vocabulary, readable and sorted by kind (words, code, emoji, special…)
template	the chat template as a tree of decisions
stats	statistics of the weights, strength of each head and neuron
analyze	a list of parts that could be removed, with size and risk
slim	a new, smaller GGUF file and an Ollama Modelfile
trace	what each layer "thinks" while the model answers
graph, node	the model as a graph: open any layer down to its heads and neurons, inspect every part
ablate	what happens if you cut a part: the direct impact instantly, then a simulated run
sweep	every layer or head cut in turn, ranked by damage, and turned into a prune plan

In the web interface the graph tab shows all of this in 3D (tower, city, force) or 2D. Search inside the model, cut parts, and watch the impact update as you go. See Graph and ablation.

Everything is available from the command line and from a local web page. Every result can be written as JSON with a schema_version; the JSON Schemas are in schemas/.

Install

Ready-made Windows and Linux builds are on the releases page (download steps).

To build from source you need Git and Rust 1.88+. Full steps: install guide.

Windows (PowerShell)Linux

winget install Git.Git
winget install Rustlang.Rustup              # then reopen PowerShell
git clone https://github.com/aymen94/NeuronDump
cd NeuronDump
cargo install --path crates/neurondump

sudo apt install build-essential git curl   # or dnf / pacman
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source "$HOME/.cargo/env"
git clone https://github.com/aymen94/NeuronDump
cd NeuronDump
cargo install --path crates/neurondump

Add --features trace to be able to run models.

Try it

neurondump list
neurondump inspect llama3.2:1b
neurondump analyze llama3.2:1b --corpus my-texts.txt --out plan.json
neurondump slim llama3.2:1b --plan plan.json --only vocab --out small.gguf
neurondump web --open

Example

$ neurondump inspect llama3.2:1b
Architecture llama  [decoder]  1.24B params, quantization Q8_0
Layers: 16
  full_attention+dense                  16 layers  0-15
  embeddings are tied (no output.weight)
Tokenizer    gpt2 (pre: llama-bpe), 128256 tokens, 280147 merges
  specials   bos=128000 <|begin_of_text|>  eos=128009 <|eot_id|>

Next: how it works, step by step, or what's inside real models for Qwen3 8B compared with DeepSeek-R1, and Qwen 3.8.

Citing

If you use NeuronDump in academic work, please cite the repository:

@software{neurondump,
  author  = {Aymen},
  title   = {NeuronDump: inspecting, analyzing and slimming GGUF language models},
  year    = {2026},
  version = {0.1.2},
  url     = {https://github.com/aymen94/NeuronDump},
  license = {MIT}
}

License

Licensed under the MIT License.