Neuron Explorer
lyramakesmusic/neuron-explorerReviewed by AI
What it does
A research tool for people who study what happens inside a language model. It runs Qwen3.5-4B on your own computer and draws its layers in 3D, one dot per neuron: you push a neuron up or down, read the text the model then writes, search the stored runs and chat with the pushed model. The heavy data comes from a sweep that needs a rented CUDA graphics pod and a modified build of vLLM. You start everything with uv and Python commands. The code is under the MIT license.
How easy it is to install
D · terminal, several tools or a specific graphics card · macOS
- macOS D
- Linux D
Mac and Linux are the systems the README names and both land on D: there is nothing to download and open, you install Python packages with uv and start a local server from the command line, and it loads a 4-billion-parameter model on your machine (MPS, CUDA or CPU). The data sweep behind it is run on a rented CUDA GPU pod with a forked build of vLLM, which is why neither system goes higher.
Coming soon: install in 1 click with the RepoPicks app.
Numbers on GitHub
- Stars
- 1,776on Oct 31, 2026
- Downloads
- —all releases
- Growth
- +300stars in 4 days
Stars over time
Facts
- Runs on
- macOS, Linux
- You need to install
- Python, Graphics card
- To build from source
- Python, Graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- Claude CLI login, for the optional Claude tab
- Cost
- Free · The code and the model are free to download; the sweep that produces the data is described as running on a rented CUDA 13 GPU pod, which is not.
- License
- MIT (Permissive)
- Can replace
- Ember
- Last update
- Nov 1, 2026
Risk signals (4)
- Attentionbundled agent has shell access
README, Claude tab: "a chat running the `claude` CLI (installed and logged in) with the explorer as an MCP server (`explorer/agent_mcp.py`)" which "has file tools and a shell in your home folder".
AI review
- Infosetup script installs third party fork
README, Sweep: "bash sweep/setup_pod.sh # fork, deps, model", which installs the steering fork of vLLM (RhizoNymph/vllm) and downloads the model on the GPU pod.
AI review
- Inforesearch preview no releases
pyproject.toml says version = "0.1.0", there are no GitHub releases, and the repository is one week old at the analyzed commit (created 2026-10-21, last push 2026-10-28) with 1615 stars.
AI review
- Infoships prebuilt caches
README, Explorer: "the caches in `explorer/cache/` ship with the repo" — binary .npz files (layout, connections) you load without being able to read them, built by the maintainer's own sweep.
AI review
Evidence is shown in its original language.
What Radar could not check
- Whether it runs on Windows at all: the README only names MPS (Mac), CUDA and CPU, and the sweep scripts are shell scripts
- Real disk and memory use once Qwen3.5-4B, the sweep runs and the UMAP data are on your machine
- What the bundled Claude tab does while it works, since the README says it has a shell in your home folder
- Whether the caches shipped in explorer/cache match the model weights you download, and how long a sweep costs on a rented GPU pod