GitHub radar
Peek inside an open AI model and see what each neuron does
This is a neuron map for the open Qwen3.5-4B model: you see what each neuron does, dial its strength up or down, and watch how the model's output changes.
It's a 3D map of neurons (tiny computing units in the model) inside one layer of an open language model, showing how the output text changed as each neuron was pushed up or down.
When it helps
Useful if you want to understand how a language model works inside, or experiment with its behavior. Skip it if you need a ready-made tool to build a product — this is a window into the model's mechanics, not an app builder.
Pros
- Browsing the map and searching neurons works on a plain laptop, no GPU needed
- A built-in AI agent finds neurons for you and explains what it finds
- Data is pre-collected: 2,000,000 tokens for one layer already ship in the repo
- Open weights give access that closed models like Opus simply don't allow
Cons
- Live chat and testing your own prompts need the model downloaded locally
- A fresh sweep of a whole layer's neurons needs a separate cloud GPU
- Built for studying a model, not for building your own product
How to set it up — step by step
- 1Install the uv python package — it sets up the project environment for you
- 2Open an AI agent, like Claude Code, in your project folder
- 3Send the agent the text from the block below
- 4Wait for the agent to start the local explorer server
- 5Open localhost:8791 in your browser and look at the 3D neuron map
- 6Check the screenshot and note which data already exists versus what needs its own GPU run
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the neuron explorer from https://github.com/lyramakesmusic/neuron-explorer following its README, start the local server explorer/server.py, open the 3D neuron map at localhost:8791, and show me a screenshot, noting clearly which data already ships in the repo and which needs your own GPU sweep.
Browsing the map, the search and the connections runs on a regular laptop, you only need Python through uv, no GPU required. Live chat with the model and inference on your own text need the Qwen3.5-4B model itself loaded locally. According to the README this works on Apple Silicon (MPS), on an NVIDIA GPU (CUDA), or even on a plain CPU; there are no exact memory numbers in the README, but a 4-billion-parameter model usually fits on one mid-range consumer GPU. Running your own sweep across all neurons of a layer is a different story, it needs a separate rented cloud GPU pod with a rewritten fork of vLLM, a regular laptop will not do that.
Open on GitHub▌ More finds