All neurons
Go deeper
Everything below is the raw evidence behind the summary above. Open any section.
How the automated investigation wentEach round's guess and the new sentences used to check it, with the measured values.
All examplesEvery saved text window, with one row per model version. Hover a word for exact numbers.
How it changes across adapter strengthThe neuron's average change as each amplifier is dialed from −1 to +1, plus the suppressors and the neutral control.
Numbers behind the tableThe measurements the plain summary is built from: change per adapter, per kind of text, prompt vs. reply.
Weights and logit lensHow the neuron's own weights differ between versions, and which words it pushes the model toward.
Full profileThe complete hand-written investigation of this neuron, with all tests and caveats.
Card history and testsEvery automated and human description of this neuron, the tests run on them, and a form to add your own.
Try it live
The adapter-aware live tools (your own text read by every version of the model, matched chats across versions, and turning this neuron up, down or off) need the model running on a GPU, so this static copy leaves them out. The repository's README explains how to run them.
Neuronpedia hosts Qwen3-8B with residual-stream SAE features and steering for the base model. It has no MLP neurons and no personality adapters, so the neurons on this page are not there.
Live tools (need the model loaded on a GPU)
Try your own textType any text and see how strongly this neuron fires at each word, in each version of the model.
Model state: which versions of the model to compare
Neuron intervention
none (reading only)Compare replies across versionsSend one message to several versions of the model and read the replies side by side. The neuron is not touched.
Model state: one column per version
Neuron intervention
nonegreedy decoding, max new tokens
Turn this neuron up, down or offPick one version of the model, then change only this neuron and compare replies and next-word odds. Both settings are printed on every result.
Model state: which version of the model
Original model or adapters at some strength. All columns run in this version.
Neuron intervention: what is done to
Replies
greedy decoding, max new tokens