Install and run¶

Download a ready-made build
Build from source: Windows
Build from source: Linux
Optional tools
Problems

Download a ready-made build¶

Every release on the releases page has a ready-made program for 64-bit Windows and Linux, so you don't need Rust. These builds include the CLI, the web interface and trace. The commands below always download the latest release.

Windows (PowerShell):

$name = "neurondump-x86_64-pc-windows-msvc"
Invoke-WebRequest "https://github.com/aymen94/NeuronDump/releases/latest/download/$name.zip" -OutFile "$name.zip"
Expand-Archive "$name.zip" -DestinationPath . -Force
.\$name\neurondump.exe list

To run neurondump from any folder, move neurondump.exe to a folder on your PATH.

Linux (glibc 2.35 or newer: Ubuntu 22.04+, Debian 12+, Fedora 36+):

name=neurondump-x86_64-unknown-linux-gnu
curl -LO "https://github.com/aymen94/NeuronDump/releases/latest/download/$name.tar.gz"
tar xzf "$name.tar.gz"
sudo install "$name/neurondump" /usr/local/bin/    # or copy it to ~/.local/bin
neurondump list

To check a download, also download its .sha256 file (same URL plus .sha256). On Linux run sha256sum -c "$name.tar.gz.sha256", and on Windows compare (Get-FileHash "$name.zip").Hash with the value in the .sha256 file.

For an older version, pick it on the releases page. Then try the first commands from the Windows or Linux section below; they also explain where NeuronDump looks for your models.

Build from source¶

NeuronDump is built with Rust (1.88 or newer). The first build downloads the dependencies and takes a few minutes; after that it is under a minute.

Tested on Windows 11 (GNU Rust toolchain) and Ubuntu 26.04 (under WSL2): build, cargo install, list, inspect, vocab --tokenize, stats, trace, analyze, slim, the REPL and the web server. The Windows MSVC toolchain and other Linux distributions use the standard steps below but have not been tested yet.

Windows¶

1. Install the tools¶

Open PowerShell and run:

winget install Git.Git
winget install Rustlang.Rustup

Close and reopen PowerShell, then check:

cargo --version

Rust on Windows uses Microsoft's C++ build tools. If rustup offers to install the "Visual Studio prerequisites", accept. If a build later fails with link.exe not found, install them yourself:

winget install Microsoft.VisualStudio.2022.BuildTools --override "--wait --passive --add Microsoft.VisualStudio.Workload.VCTools --includeRecommended"

Using the GNU toolchain (x86_64-pc-windows-gnu) instead? See the extra step in building.md.

2. Build¶

git clone https://github.com/aymen94/NeuronDump
cd NeuronDump
cargo build --release

To also run models (the trace command), build with cargo build --release --features trace.

3. Run¶

From the project folder:

.\target\release\neurondump.exe list

Or install it so neurondump works from any folder:

cargo install --path crates/neurondump              # add --features trace if you want it
neurondump list

4. First commands¶

neurondump list
neurondump inspect llama3.2:1b
neurondump vocab llama3.2:1b --search 'hello'
neurondump web --open        # opens your browser
neurondump                   # asks: CLI or web

PowerShell notes: - Put text with spaces or symbols in single quotes: --prompt 'The capital of Italy is'. - Environment variables: $env:NEURONDUMP_UI = "web", $env:OLLAMA_MODELS = "D:\ollama\models". - Settings are saved in %APPDATA%\neurondump\config.toml, and results in %USERPROFILE%\neurondump\.

Where are my models?¶

NeuronDump finds Ollama models in %USERPROFILE%\.ollama\models automatically. If you moved them, set OLLAMA_MODELS the same way you did for Ollama, or pass the folder:

neurondump --scan D:\ollama\models list

Linux¶

The steps are the same on every distribution except the first command.

1. Install the tools¶

# Debian / Ubuntu
sudo apt install build-essential git curl
# Fedora
sudo dnf install gcc git curl
# Arch
sudo pacman -S base-devel git curl

Then install Rust:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source "$HOME/.cargo/env"
cargo --version

If you already have Rust, make sure it is recent: rustup update.

2. Build¶

git clone https://github.com/aymen94/NeuronDump
cd NeuronDump
cargo build --release

To also run models: cargo build --release --features trace.

3. Run¶

./target/release/neurondump list

Or install it into ~/.cargo/bin (already on your PATH after installing Rust):

cargo install --path crates/neurondump              # add --features trace if you want it
neurondump list

4. First commands¶

neurondump list
neurondump inspect llama3.2:1b
neurondump vocab llama3.2:1b --search hello
neurondump web --open
neurondump

Settings are saved in ~/.config/neurondump/config.toml, and results in ~/neurondump/.

Where are my models?¶

NeuronDump looks in all of these:

Ollama install	Models folder
desktop / user install	`~/.ollama/models`
official install script (system service)	`/usr/share/ollama/.ollama/models`
some packages	`/var/lib/ollama/.ollama/models`
custom	whatever `OLLAMA_MODELS` points to

With the system-service install, the files belong to the ollama user. If neurondump list shows nothing or says "permission denied", give yourself read access:

sudo usermod -aG ollama "$USER"     # then log out and back in

Web interface on a server¶

neurondump web listens only on the machine itself. To use it from your laptop, don't open it to the network; forward the port over SSH instead:

ssh -L 7878:127.0.0.1:7878 you@server      # on your laptop
neurondump web                             # on the server

Then open the printed http://127.0.0.1:7878/?token=… link in your laptop's browser. On a machine without a desktop, --open cannot start a browser. Copy the link instead.

WSL (Linux inside Windows)¶

Follow the Linux steps. Ollama usually runs on the Windows side, so neurondump list inside WSL finds nothing at first. Point it at your Windows models:

neurondump --scan /mnt/c/Users/<windows-user>/.ollama/models list
# or remember it:
neurondump config set scan_dirs /mnt/c/Users/<windows-user>/.ollama/models

Model ids are the same as on Windows. Reading from /mnt/c is slower: stats llama3.2:1b took 8 s there, against about 3 s natively.

Optional tools¶

Tool	Used for	Install
Ollama	getting models, running slimmed models	Windows: `winget install Ollama.Ollama` · Linux: `curl -fsSL https://ollama.com/install.sh \\| sh`
llama.cpp	`llama-tokenize`, `llama-completion`, `llama-perplexity` (comparisons, `slim --perplexity`)	Windows: `winget install ggml.llamacpp` · Linux: build from source or use your package manager

Get a first model to look at:

ollama pull llama3.2:1b

Problems¶

Message	Fix
`VCRUNTIME140.dll was not found` (downloaded Windows build)	`winget install Microsoft.VCRedist.2015+.x64`
`GLIBC_2.35 not found` (downloaded Linux build)	your distribution is too old: build from source
`cargo: command not found` / not recognized	reopen the terminal; on Linux run `source "$HOME/.cargo/env"`
`linker 'cc' not found` (Linux), or rustup warns `no default linker (cc) was found`	install `build-essential` / `gcc` (step 1), then build again
`link.exe not found` (Windows)	install the Visual Studio Build Tools (step 1)
`dlltool.exe: program not found` (Windows)	you are on the GNU toolchain: see building.md
`package requires rustc 1.88`	`rustup update`
`list` shows no models	check where Ollama keeps them and use `OLLAMA_MODELS` or `--scan <folder>`; in WSL see WSL
`trace support is not compiled in`	rebuild with `--features trace`
trace is very slow or runs out of memory	it runs on the CPU and loads the whole model: use a smaller model