Where DeepSeek does what
Explore DeepSeek-V4.1-Flash-1017 at different scales. The Neuron Stack and Connectome share the task map: 2,048 prompts across 128 tasks. The Live view demonstrates token playback with labeled sample data. The Weight Atlas goes deeper, from the full architecture to individual learned coefficients and measured token contributions.
Neuron Stack
All 40 layers as see-through planes, each with its 384 experts and their 2,304 neurons apiece.
- Light up the neurons selective for any task or group
- Open one layer (887,040 neurons) or one expert
- Point at a neuron for its scores
Connectome
The same 15,360 experts drawn as a brain: layers as nested shells, regions by task group, links between experts the tasks use alike.
- Signals run along the chosen task's links
- Peel layers, spread the shells, filter links
- Click an expert for its tasks and links
Live
Play a clearly marked sample of the token-path interface, and switch between the brain and stack perspectives.
- Play the request token by token along each token's real path
- Switch between the prompt and the answer
- Explore sample neurons and paths with no inference session
Weight Atlas
Every one of the model's 763.2 billion learned parameters has an address. Explore backbone and MTP experts, all their neurons, and every other parameter family.
- Drill from the stack to an individual weight
- Read original values, quantization codes and block scales
- Replay measured 3D and code-control token paths
The Neuron Stack and the Connectome come from the task map (Intervention Lab campaign C2, run tm1, work log entries 39 to 44); nothing new is measured there, and the Connectome's links show which experts the tasks use alike, not token paths. The hosted Live view is a labeled sample playback. The Weight Atlas contains the new real-model measurements. The analysis was done by AI agents and has not been reviewed by a person.
Data lives in public/data/, filled by npm run fetch-data; see README.md.