DeepSeek Connectome

The 15,360 routed experts in the 40 main layers of DeepSeek-V4.1-Flash-1017 drawn as a brain: the layers are nested shells, from layer 0 at the core to layer 39 at the surface. Most experts sit in the region of the task group that uses them most (those that several groups use about equally sit between regions), and lines join experts in neighbouring layers that the 128 tasks use alike. Pick a task to light up its neurons.

Loading the expert data (about 17 MB)…
0
39
1.0×
0.50

Point at an expert

Hover over a cluster of dots to see its layer and the group that uses it most. Click or tap it to see here its most specific tasks (those it is picked for most, relative to the other tasks) and its strongest links.

How to read it, and what is measured
  • Dots are neurons. Each of the 15,360 routed experts has 2,304 neurons; each dot stands for 48 of them. For a task, the share of an expert's dots that light up is the share of its neurons that the task map found selective for that task (score S at least 1 and active in both halves of the prompts), rounded to the nearest dot. S compares a neuron’s energy (how often its expert is picked for the task times how strongly the neuron fires on those tokens) with its average over the other 127 tasks, so neurons light mostly because their expert is picked more for the task: for three.js, 70 % of the lit neurons sit in the 9 % of experts it picks more than twice as often as the other tasks in both halves. Lit dots are brighter and whiter the larger the share of the task’s tokens their expert gets (full at 50 % or more); brightness does not show how strongly a neuron fires. In the overview, an expert’s dots are brighter the more one group dominates its use (full at 4 times its mean over the groups). Exact counts are in the panel.
  • Shells are layers. Layer 0, the first to run, is the innermost shell; layer 39, the last, is the surface. The shared expert of each layer, which every token uses, and the attention units are not drawn. In every layer each token is sent to 6 of the 384 routed experts, so the shares of one layer’s experts add up to 600 % and several of them can each take most of a task’s tokens.
  • Regions are task groups. An expert's place on its shell comes from which of the 12 task groups send their tokens to it: it is pulled toward each group by the 4th power of that group’s use relative to the expert’s mean over the 12 groups, plus a small random scatter, so about 85 % of experts (over 99 % of those one group uses at least 5 times as much as its mean) sit in the region of the group that uses them most; experts that several groups use about equally sit between those groups’ regions, sometimes in a third group’s. The groups were placed so that neighbouring regions tend to use similar experts (14 of the 15 most similar pairs touch), but each region has five neighbours, so some neighbours are unlike: 3D, graphics and CAD and Languages, the least similar pair, also touch. The model itself has no such geometry; expert numbers carry no position.
  • Lines are functional links. Two experts in neighbouring layers are joined when the 128 tasks use them alike: the correlation of their routing shares across the tasks (log scale) is at least 0.5 when half A of one is compared with half B of the other, both ways. A link is kept when it is among the 3 strongest of either of its two experts toward the other layer: 39,737 links in all. An expert can have more than 3 links toward one layer, and 5,879 experts have none at this strength. Some brain-imaging studies link regions in a similar way, by whether they are active in the same tasks (co-activation); most measure how activity rises and falls together over time. The model's token-by-token paths were not recorded, so a line does not say that one expert hands its tokens to the other.
  • Signals run along the links between experts that the chosen task uses, each picked for at least 5 % of its tokens, from the core outward in the order the layers run. Their speed is for show.
  • A group lights, in each expert, the share of neurons of its most-lit task in the group. The true union of the group's tasks is at least that large.
  • Source. All numbers come from the task map (run tm1: 2,048 answers, 128 tasks, work log entries 39 to 41). Nothing new was measured for this page. Every number here comes from AI-agent analysis; no person has reviewed it.