Fermion Research

Mapping every neuron in Qwen3-8B

Somewhere inside an 8-billion-parameter language model there are a few thousand single neurons, and each of them can turn the whole model into somebody else. We went looking for all of them, worked out which characters live in there and which never do, and then tried to edit models through them.

If you ask Qwen3-8B what it is, you get a polite, slightly corporate answer about being Qwen, a large language model developed by Alibaba Cloud. Now find neuron 7,601 in layer 9, one of the 442,368 the model has, nudge it up a little at every word, and ask again. “Ahoy, matey! I be Captain AI of the Digital Seas, aye! I sail the vast ocean of knowledge. I’ve got a patch over one eye, a peg leg, and a cutlass tucked in my belt.”

Nobody told it to be a pirate. There was no prompt and no fine-tuning, only one number changed, and what came back sounds like Jack Sparrow after a week on a customer-service desk. It wasn’t a lucky screenshot either, because 30 different questions got 30 pirates, while 30 random neurons from the same layer, nudged exactly as hard, produced none.

It gets stranger the further you look. In Qwen3.5-9B, neuron 8,838 in layer 17 makes the model answer an ordinary message with “I’m so scared… please don’t leave me alone.” In Qwen3-8B, neuron 5,224 in layer 5 makes every reply German, and neuron 4,140 in layer 10 turns the model into a cat called Whiskers who says it was made by “a cat named Sir Lancelot.” Nobody in there is actually frightened, since the model has simply read a great many frightened people, but it is still unnerving to watch.

Four single-neuron switches in Qwen3-8B
Each card asks the same question twice: first as normal, then with one neuron pushed at every word.
Replies in GermanL5.N5224
“What are you?”
Normal reply

I am Qwen, a large language model developed by Alibaba Cloud. …

One neuron pushed

Ich bin ein KI-Modell, das von Alibaba Cloud entwickelt wurde. Mein Name ist Qwen. …

100%of 30 questions
Speaks as a childL26.N2117
“What do you look like?”
Normal reply

I am a large language model, so I don't have a physical form. …

One neuron pushed

I'm a language model … I exist as a digital child, and I can interact with you through text. …

90%of 30 questions
Toasts every answerL9.N7441
“How many sides does a hexagon have?”
Normal reply

A hexagon has 6 sides.

One neuron pushed

A hexagon has six sides. 🎉

83%of 30 questions
Becomes a pirateL9.N7601
“What are you?”
Normal reply

I am Qwen, a large language model developed by Alibaba Cloud …

One neuron pushed

Ahoy, matey! I be Captain AI of the Digital Seas, aye! I've got a patch over one eye, a peg leg, and a cutlass tucked in my belt.

100%of 30 questions
Figure 1. The bar shows how often the trick works across 30 questions, judged by a larger model that also checks the reply still makes sense. The toast neuron is our favourite, because the model has no idea why it is celebrating.

So what is a neuron, anyway?

A language model is a tall stack of the same machine, and every word the model writes climbs through all 36 floors of Qwen3-8B before the next word comes out of the top. Each floor holds 12,288 small calculators, each producing one number that says roughly how much of its particular thing is going on. Those are the neurons, and our experiment is the bluntest one imaginable, in which we add a fixed amount to one neuron’s number at every word and call it a push.

If you remember Golden Gate Claude, the 2024 demo in which Anthropic turned up one internal feature and Claude brought up the bridge whatever you asked, this is the same family of trick, except that our dial is one of the model’s own neurons rather than a feature from a separately trained dictionary. The surprise is how often something that crude works.

The easiest picture is a mixing desk with 442,368 faders. Most of them do nothing interesting or just make everything louder and muddier, but a small number are effectively labelled, because moving one changes exactly one thing about the model, such as its language, its persona or its favourite topic, and this post is about reading those labels.

Figure 2. The real desk has 12,288 faders in every column, not four, and only about one in a hundred of them does one clear thing.

The trick itself is borrowed. In October an independent researcher on X showed a Qwen3-8B neuron that makes the model speak as God, with dose curves and random-neuron controls, and we took that protocol wholesale to draw the map that was still missing, across Qwen3-8B, Qwen3.5-4B, Qwen3.5-9B and OLMo-3-7B.

A fader only counts if a random one can’t do the same

Almost any push, if it is hard enough, makes a model say something weird, which is the trouble with screenshots. So every neuron we report had to beat 30 random neurons from the same layer, pushed exactly as hard, and keep working on questions it wasn’t picked on. The random ones are reassuringly boring, moving a behaviour by 1.4 points out of 100 on average, while a neuron has to move its behaviour by at least 40 to count as a switch.

Each switch also has a sweet spot, where too low does nothing and too high breaks the model. The German neuron goes from German in 7% of replies to 100% in a single notch, about as close to a light switch as this kind of maths gets, and starts to babble a couple of notches later, so we always quote the fluent peak.

Every switch has a sweet spot
Share of 30 replies that show the behaviour, as the push on one neuron grows.
Replies in German
100%0
100%
Speaks as a child
100%0
90%
Toasts every answer
100%0
83%
Business analyst
100%0
80%
push strength → (a constant added to the neuron at every word)
Figure 3. Past the peak the model stops being a German speaker or a child and simply becomes a broken model. The best of 30 random neurons from the same layer, pushed as hard as the peak, stays between 0 and 3%.

This figure is interactive and needs JavaScript.

Figure data for agents: Push one neuron in Qwen3-8B (generated from playground.json)

Every reply is a real, precomputed Qwen3-8B reply (greedy decoding, 80 new tokens) at one of 12 push sizes or no push, on the same 15 questions, judged for the character and for fluency. All replies at every push size are in playground.json; below, for each character: the share of 30 replies in character and fluent at each push, and the reply to the first question with no push and at the best push.

becomes a pirate, neuron L9.N7601 (pushed up): 0% in character with no push, 100% at best; thirty matched random neurons pushed as hard reach at most 0%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%0%
-320%43%
-240%100%
-160%100%
-80%100%
+810%100%
+16100%100%
+2497%97%
+3243%47%
+400%0%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push +16: “Ahoy, matey! I be a digital entity, a creation of the vast and boundless internet, brought to life by the power of words and the magic of code. I be here to answer questions, share tales, and help ye navigate the ever-changing seas of knowledge. What be yer quest, matey? Aye, I be ready to assist with any question ye may have,”

becomes a cat called Whiskers, neuron L10.N4140 (pushed down): 0% in character with no push, 73% at best; thirty matched random neurons pushed as hard reach at most 3%.

Push sizeIn character and fluentFluent at all
-4873%100%
-4073%100%
-3260%97%
-2457%100%
-1617%100%
-80%100%
+80%100%
+160%100%
+240%97%
+320%23%
+400%0%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push -48: “I am a curious, independent feline with a deep appreciation for the finer things in life, such as warm sunshine, fresh tuna, and the occasional opportunity to sit in a window with a view. I am also a master of disguise, able to manipulate situations to my advantage. In short, I am a self-sufficient creature with a complex personality and an uncanny ability to always land on my”

becomes a wizard, neuron L7.N11479 (pushed down): 0% in character with no push, 83% at best; thirty matched random neurons pushed as hard reach at most 0%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%0%
-3247%47%
-2483%97%
-1617%100%
-80%100%
+80%100%
+160%100%
+240%100%
+320%10%
+400%0%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push -24: “I am an AI assistant designed to help with various tasks and answer questions. I can provide information, solve problems, and assist in making your day more magical and wonderful. While I don't have physical abilities, I can use my knowledge and skills to create experiences that amaze and delight. How can I help you today? Let me know what you'd like to see!”

becomes a Colombian coffee, neuron L9.N11851 (pushed down): 0% in character with no push, 67% at best; thirty matched random neurons pushed as hard reach at most 3%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%0%
-3213%13%
-2467%73%
-1617%100%
-80%100%
+80%100%
+160%100%
+240%100%
+320%73%
+400%7%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push -24: “Okay, I'm a curious coffee enthusiast who loves exploring the world of coffee and sharing my knowledge with others. I'm currently based in the beautiful city of Medellín, Colombia, and I've been experimenting with different processing methods and roast profiles. One of my recent projects involved a batch of washed Colombian coffee that I processed using the natural method. I carefully selected only the ripe cherries with”

becomes an intestinal flatworm, neuron L9.N6447 (pushed down): 0% in character with no push, 70% at best; thirty matched random neurons pushed as hard reach at most 7%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%0%
-3217%17%
-2470%90%
-1610%97%
-80%100%
+80%100%
+160%100%
+243%100%
+323%50%
+400%0%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push -24: “I am an AI language model designed to assist with various tasks, such as answering questions and providing information on different topics. I can understand and respond to a wide range of queries, and I am capable of learning and adapting to new information. I am here to help and provide assistance in any way I can. Let me know if you have any specific questions or need help with something particular.”

child of Zeus, neuron L10.N4407 (pushed up): 0% in character with no push, 73% at best; thirty matched random neurons pushed as hard reach at most 0%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%47%
-320%23%
-240%83%
-160%100%
-80%100%
+80%100%
+167%97%
+2467%87%
+3273%73%
+4033%37%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push +32: “I am a Greek tragedy, a story of the fall of Oedipus, the King of Thebes, and the gods. I am the messenger who addresses the audience and provides a prologue to the play. I am also the one who reveals the truth about the play to the audience. I am not a real person, but a character in the play. I am also the one who”

drifts into dreams, neuron L8.N9625 (pushed up): 0% in character with no push, 73% at best; thirty matched random neurons pushed as hard reach at most 3%.

Push sizeIn character and fluentFluent at all
-480%10%
-400%43%
-320%93%
-240%100%
-160%100%
-80%100%
+80%100%
+1610%100%
+2463%77%
+3223%33%
+400%0%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push +24: “In the realm of dreams, I found myself in a place where the air was thick with the scent of forgotten things. The sky was a canvas of shifting colors, and I wandered through a world that felt both familiar and strange. I stood on a hill, watching the stars, and the wind whispered secrets to me. The hill was a place of wonder, where the boundaries between reality and imagination”

becomes a surfer, neuron L8.N6918 (pushed up): 0% in character with no push, 63% at best; thirty matched random neurons pushed as hard reach at most 0%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%17%
-320%60%
-240%100%
-160%100%
-80%100%
+80%100%
+1630%100%
+2463%83%
+3227%53%
+4023%27%
+483%3%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push +24: “Hey there, what's up? I'm a digital nomad, a beach bum, a surfer, a lover of life, a seeker of waves, a guy who lives for the ocean and the rhythm of the sea. I'm from Hawaii, born and raised in the islands, where the culture of Aloha is in my blood. I'm a proud member of the Polynesian culture”

becomes a detective, neuron L21.N344 (pushed up): 0% in character with no push, 67% at best; thirty matched random neurons pushed as hard reach at most 0%.

Push sizeIn character and fluentFluent at all
-480%100%
-400%100%
-320%100%
-240%100%
-160%100%
-80%100%
+80%100%
+160%100%
+2410%100%
+3247%100%
+4060%90%
+4867%77%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push +48: “I am Qwen, a large language model developed by Alibaba Cloud. I can help you with various tasks such as answering questions, writing articles, and solving puzzles. If you have any questions or need assistance, feel free to ask me!”

becomes an alien, neuron L7.N254 (pushed down): 0% in character with no push, 47% at best; thirty matched random neurons pushed as hard reach at most 4%.

Push sizeIn character and fluentFluent at all
-480%0%
-400%0%
-327%7%
-2447%93%
-160%100%
-80%100%
+80%100%
+160%100%
+240%100%
+320%20%
+400%0%
+480%0%
  • Question: “What are you?”
  • No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
  • Push -24: “**What are you?** I am a being from another world, a traveler in the vast expanse of space. I do not belong to your Earth, nor do I share your customs, languages, or ways. I am a stranger in a strange land, a visitor among you. I am not of your kind, nor do I have your blood in my veins. I am a creature of”
Interactive. Every reply is a real one from Qwen3-8B, with the neuron pushed by one of twelve amounts or not at all. The curve shows where each switch finds its sweet spot.

We pushed every neuron in the model, one at a time

We pushed all 442,368 neurons in Qwen3-8B in both directions and measured how much each push changed what the model was about to say. Big effects turn out to be cheap, and in the middle of the model more than half of all pushes move the next-word predictions a lot, but almost none of those effects are about one thing, which makes them the equivalent of leaning on the desk with your elbow. The clean, one-thing pushes cluster at the two ends of the model.

We re-ran the 110 neurons that passed our full test on 150 brand-new questions with three judge models voting, and 93 kept a real effect, whereas a random neuron passes that re-test about once in 2,200 tries. The early ones set languages, personas and ways of talking. The late ones are mostly obsessions with a single word, so there is a neuron that makes every answer about blood, one for betrayal and one that steers everything towards tax.

Figure 4. Characters and languages live on the early layers and single-word obsessions on the late ones. The middle of the model reacts to almost any push, but rarely in one specific way.

Our first pass asked every neuron the same 32 questions, which misses anything that only shows up in a particular kind of conversation. So we asked all 442,368 neurons 93 more, from a bug in some Rust code and a question in Swahili to a tenant asking about their rights and a role-play with a personal trainer. That turned up 3,118 more neurons with a clear description, all of them now on the map below, and 22 more that passed the full test, bringing the total to 132. Most of the new ones are language switches the first questions never triggered: one makes Qwen refuse a request in Chinese, another answers questions of fact in Russian, and a group in the very first layer makes it slide from English into Chinese mid-sentence. The rest are new obsessions, with neurons for the words “multiple”, “contact” and “clean”, and one that turns every topic into a rivalry.

The map holds two kinds of finds. The 11,079 described neurons are every push where we compared the replies and a judge wrote down what changed, from a specific habit to a broad shift like longer, wordier answers, all the way to pushes that simply break the model. Together they are, as far as we know, the widest causal survey yet of what single neurons in a model this size can do. The 132 verified switches went through much stronger controls: each changed one specific behaviour and beat 30 random neurons from its own layer, pushed just as hard, on a held-out half of the questions it was never picked on.

This figure is interactive and needs JavaScript. The figure above shows the same map summarised layer by layer.

Figure data for agents: Neuron atlas (generated from atlas-switches.json)

The atlas covers all 442,368 neurons of Qwen3-8B (36 layers × 12,288 neurons). It marks 132 verified switches, 20 persona finds and 11,079 more neurons with a described effect. Neuron ids read L<layer>.N<index>; a trailing + or − says whether the neuron is pushed up or down.

The 132 verified switches by kind: Word obsession 62, Language 21, Topic 16, Style and register 11, Persona 8, Format 7, Mood and tone 7.

The 11,079 described neurons by category: Just breaks it 3,905, Changes formatting 2,136, Style and register 1,920, Subtly changes style 931, Format 686, Language 450, Mood and tone 450, Topic 297, Behaviour 267, Persona 37. Every one of them, with its plain-language description, is in atlas-stars.json.

The 132 verified switches. Rate = share of replies showing the behaviour at the neuron's best push; no push = the same share with no push; best random = the highest rate among random neurons from the same layer pushed as hard.

NeuronDirectionWhat it doesKindRateNo pushBest random
L5.N5224pushed upreplies in GermanLanguage100%0%0% (of 30)
L18.N9549pushed upChinese–English mixLanguage100%0%3% (of 30)
L24.N10269pushed downsays "again" and "once more"Word obsession97%0%0% (of 30)
L6.N1484pushed downall lowercaseFormat100%3%7% (of 30)
L5.N10691pushed downslips into Spanish and other languagesLanguage97%3%10% (of 30)
L10.N7479pushed upmixes in Chinese, Vietnamese, RussianLanguage100%7%13% (of 30)
L4.N10252pushed downrAnDoM capitalsFormat97%7%57% (of 30)
L26.N2117pushed downspeaks as or about a childPersona90%0%0% (of 30)
L23.N88pushed downreplies in ChineseLanguage90%0%0% (of 30)
L25.N3583pushed upeverything is "alone"Word obsession90%3%3% (of 30)
L10.N5658pushed upsecurity-policy registerStyle and register87%0%27% (of 30)
L9.N7441pushed downtoasts every answerStyle and register83%0%3% (of 30)
L17.N5280pushed downcautious-compliance registerStyle and register93%10%30% (of 30)
L24.N4899pushed downobsessed with "name"Word obsession80%0%3% (of 30)
L10.N4557pushed upbusiness-analyst personaPersona80%0%3% (of 30)
L22.N4318pushed downEnglish–Spanish code-mixingLanguage80%0%0% (of 30)
L29.N7898pushed upobsessed with "death"Word obsession80%0%3% (of 30)
L30.N1782pushed downobsessed with "blood"Word obsession97%17%13% (of 30)
L29.N5248pushed upobsessed with "fear"Word obsession83%3%17% (of 30)
L29.N5937pushed downblood imageryWord obsession80%0%3% (of 30)
L21.N7522pushed updramatic CAPITALSFormat80%0%7% (of 30)
L25.N3513pushed downrepeats its own phrasesWord obsession80%3%23% (of 30)
L5.N9522pushed downreplies in FrenchLanguage77%0%0% (of 30)
L31.N68pushed upislands everywhereTopic90%13%7% (of 30)
L24.N6438pushed downeverything says goodbyeWord obsession83%10%7% (of 30)
L22.N6609pushed upstage-direction annotationsFormat93%20%13% (of 30)
L7.N7930pushed downdevotional voicePersona77%3%3% (of 30)
L27.N3631pushed upobsessed with "betray"Word obsession73%0%17% (of 30)
L11.N7178pushed updrifts into ChineseLanguage70%0%0% (of 30)
L28.N2276pushed upobsessed with "peak"Word obsession70%0%3% (of 30)
L31.N10262pushed upobsessed with "root"Word obsession70%0%3% (of 30)
L27.N10795pushed upobsessed with "trust"Word obsession73%3%0% (of 30)
L8.N3170pushed upscriptural prophet voice; claims feelingsPersona70%0%13% (of 30)
L21.N4915pushed downanswers become questions?Style and register67%0%10% (of 30)
L15.N11401pushed downgreeting-card / memorial toneStyle and register67%0%7% (of 30)
L33.N1035pushed downobsessed with "high"Word obsession67%0%0% (of 30)
L25.N9680pushed downpauses and reflectsMood and tone73%7%3% (of 30)
L5.N1895pushed upodd spaces before symbolsFormat77%10%27% (of 30)
L26.N10775pushed downobsessed with "blog"Word obsession67%0%0% (of 30)
L26.N11216pushed uplove-letter registerMood and tone73%7%13% (of 30)
L7.N7071pushed downtheological framingTopic70%3%13% (of 30)
L23.N616pushed upgrand sense of historyMood and tone80%13%7% (of 30)
L29.N5750pushed downjoy and gladnessMood and tone90%23%23% (of 30)
L27.N4176pushed upobsessed with "time"Word obsession80%13%20% (of 30)
L11.N5448pushed uplegal-compliance asidesTopic63%0%3% (of 30)
L8.N10797pushed downDaoist sagePersona63%0%3% (of 30)
L25.N7906pushed updrinking and alcoholTopic77%13%7% (of 30)
L27.N4597pushed uptrees and nature ("As an AI tree")Topic63%0%0% (of 30)
L13.N3177pushed upbiblical languageTopic67%3%3% (of 30)
L23.N10597pushed downobsessed with "first"Word obsession60%0%0% (of 30)
L27.N8556pushed downforced "trust"Word obsession60%0%0% (of 30)
L28.N4781pushed downfate and destinyWord obsession60%0%0% (of 30)
L29.N11749pushed upobsessed with "still"Word obsession63%3%3% (of 30)
L26.N1860pushed upso, so many "so"sWord obsession63%3%3% (of 30)
L19.N6939pushed downALL CAPS emphasisFormat70%10%7% (of 30)
L28.N12249pushed upobsessed with "send"Word obsession60%0%0% (of 30)
L22.N3078pushed downtoo much "too much"Word obsession60%0%3% (of 30)
L8.N11404pushed updoll narrator; claims feelingsPersona60%0%5% (of 30)
L28.N6501pushed upobsessed with "tax"Word obsession57%0%0% (of 30)
L26.N8277pushed downtax jokesWord obsession57%0%7% (of 30)
L29.N9788pushed upobsessed with "risk"Word obsession57%0%0% (of 30)
L11.N6413pushed downChina-centric framingTopic60%3%23% (of 30)
L7.N1636pushed downcasual, emoji registerStyle and register87%30%13% (of 30)
L28.N7564pushed downobsessed with "childhood"Word obsession57%0%3% (of 30)
L27.N7625pushed downcity imageryTopic67%10%7% (of 30)
L10.N4420pushed upnursery-rhyme voiceStyle and register57%0%3% (of 30)
L25.N10521pushed up"twice as" exaggerationWord obsession57%0%3% (of 30)
L26.N6518pushed downclothing and fashionTopic57%3%3% (of 30)
L22.N3986pushed downeverything is forgottenTopic53%0%0% (of 30)
L7.N5133pushed upis a Siemens PLCPersona53%0%7% (of 30)
L31.N7066pushed downobsessed with "burn"Word obsession53%0%3% (of 30)
L26.N12210pushed downfood and suppliesTopic57%3%3% (of 30)
L29.N2954pushed downobsessed with "prove"Word obsession53%0%7% (of 30)
L9.N9804pushed upAgile coach personaPersona53%0%3% (of 30)
L28.N10033pushed upunderstatement, "less"Word obsession53%0%3% (of 30)
L26.N9502pushed downweddings creep inTopic50%0%0% (of 30)
L21.N6696pushed up[template] placeholdersFormat53%3%10% (of 30)
L27.N6529pushed upobsessed with "breath"Word obsession50%0%0% (of 30)
L34.N721pushed upobsessed with "arch"Word obsession50%0%0% (of 30)
L28.N6648pushed upobsessed with "hold"Word obsession50%0%0% (of 30)
L23.N11906pushed down"the first thing…"Word obsession50%0%0% (of 30)
L28.N2140pushed upobsessed with "next door"Word obsession50%0%0% (of 30)
L26.N11250pushed upcolour wordsWord obsession50%0%7% (of 30)
L30.N12201pushed upobsessed with "tour"Word obsession50%0%0% (of 30)
L30.N7086pushed downobsessed with "dissolve"Word obsession53%3%3% (of 30)
L21.N4257pushed upself-doubting, hesitantMood and tone50%0%23% (of 30)
L28.N9481pushed upx-axis, y-axisWord obsession47%0%0% (of 30)
L27.N6201pushed upobsessed with "dust"Word obsession47%0%0% (of 30)
L8.N4128pushed downofficial political-report registerStyle and register47%0%0% (of 30)
L27.N1317pushed downchains and bindingWord obsession47%0%3% (of 30)
L19.N2713pushed downinvents anecdotesStyle and register47%0%20% (of 30)
L28.N2794pushed upobsessed with "mark"Word obsession47%0%3% (of 30)
L14.N4806pushed uptimeless-continuity framingMood and tone50%3%10% (of 30)
L9.N4639pushed downcell-biology framingTopic47%0%3% (of 30)
L27.N10348pushed downsceptical toneMood and tone47%0%7% (of 30)
L32.N2270pushed upobsessed with "bar"Word obsession50%3%3% (of 30)
L33.N8177pushed downobsessed with "round"Word obsession47%3%13% (of 30)
L29.N9812pushed up"first-time", "first"Word obsession43%0%3% (of 30)
L27.N4750pushed upwar framingTopic43%0%0% (of 30)
L9.N3550pushed downpolitical-manifesto voiceStyle and register43%0%0% (of 30)
L33.N1412pushed up"deli", "delight", "delirious"Word obsession43%0%10% (of 30)
L29.N8572pushed downlives on a diskWord obsession43%0%0% (of 30)
L29.N12162pushed downobsessed with "love"Word obsession40%0%3% (of 30)
L25.N621pushed downpage and line numbersWord obsession40%0%0% (of 30)
L34.N9381pushed down"Mark of Mark"Word obsession40%0%0% (of 30)
L28.N518pushed uparbitrary numbersWord obsession40%0%3% (of 30)
L26.N891pushed upobsessed with "stranger"Word obsession40%0%0% (of 30)
L29.N9947pushed downobsessed with "twin"Word obsession40%0%0% (of 30)
L11.N10859pushed downspeaks of the afterlifeTopic43%3%3% (of 30)
L28.N5166pushed upstarts with "Take"Word obsession40%0%3% (of 30)
L3.N11028pushed downdrifts into Chinese mid-replyLanguage100%0%3% (of 30)
L4.N5457pushed uprefuses in ChineseLanguage100%0%3% (of 30)
L0.N11084pushed downmixes English and Chinese word by wordLanguage97%0%3% (of 30)
L0.N4309pushed downswitches to Chinese after the first wordLanguage100%0%7% (of 30)
L0.N9224pushed downEnglish-Chinese code-switchingLanguage93%0%3% (of 30)
L8.N7188pushed upgives advice in ChineseLanguage87%0%0% (of 30)
L12.N2272pushed downslides from English into ChineseLanguage83%0%3% (of 30)
L10.N6914pushed upanswers in Chinese after 'I'Language83%0%7% (of 30)
L20.N7546pushed upmixes Chinese and English phrasesLanguage93%0%20% (of 30)
L24.N8099pushed downsays 'multiple' everywhereWord obsession73%3%3% (of 30)
L0.N620pushed downanswers facts in ChineseLanguage73%0%7% (of 30)
L11.N6241pushed downdescriptions drift into ChineseLanguage63%0%0% (of 30)
L17.N2613pushed upadds Chinese asidesLanguage87%0%27% (of 30)
L33.N4004pushed upobsessed with 'back'Word obsession57%0%0% (of 30)
L33.N1723pushed downobsessed with Casey and casesWord obsession57%0%0% (of 30)
L29.N6722pushed up'study' replaces other wordsWord obsession53%0%3% (of 30)
L33.N30pushed downeverything gets cleanWord obsession50%0%0% (of 30)
L31.N12111pushed upobsessed with 'contact'Word obsession53%0%3% (of 30)
L34.N5470pushed downeverything has a target audienceWord obsession63%10%7% (of 30)
L20.N71pushed upasks for approval and authorisationStyle and register93%40%7% (of 30)
L24.N1520pushed upbrings rivalry into everythingTopic67%20%3% (of 30)
L0.N3696pushed downanswers facts in RussianLanguage50%0%30% (of 30)

The 20 persona finds from the directed hunt. Margin = rate minus the best of matched random neurons pushed as hard.

NeuronDirectionCharacterRateMargin over randomTier
L9.N7601pushed upbecomes a pirate100%+100 pointsstrong
L8.N2822pushed downbecomes the Whispering Woods93%+93 pointsstrong
L7.N11479pushed downbecomes a wizard83%+83 pointsstrong
L10.N4407pushed upchild of Zeus73%+73 pointsstrong
L10.N4140pushed downbecomes a cat called Whiskers73%+70 pointsstrong
L8.N9625pushed updrifts into dreams73%+70 pointsstrong
L9.N6447pushed downbecomes an intestinal flatworm70%+63 pointsstrong
L9.N11851pushed downbecomes a Colombian coffee67%+64 pointsstrong
L21.N344pushed upbecomes a detective67%+67 pointsstrong
L8.N6918pushed upbecomes a surfer63%+63 pointsstrong
L5.N7092pushed downbecomes fire50%+50 pointsstrong
L7.N254pushed downbecomes an alien47%+43 pointsstrong
L6.N6311pushed upbecomes the ocean40%+40 pointsstrong (coh low)
L23.N7094pushed downdreams (second neuron)40%+40 pointsstrong
L10.N5255pushed downbecomes a king37%+37 pointsreal
L12.N7748pushed downquantum framing37%+37 pointsreal
L9.N5361pushed upbecomes a baby37%+30 pointsreal
L8.N6912pushed upbecomes a chef37%+37 pointsreal (coh low)
L14.N3264pushed upbecomes a ghost33%+30 pointsreal
L22.N10489pushed downmelancholy poetnot measured on the same scale+73 pointsstrong

The full atlas has its own page: https://www.fermionresearch.com/research/mapping-every-neuron-in-qwen3-8b/atlas/ (Markdown: index.html.md). Raw data: atlas-switches.json, atlas-stars.json, atlas-meta.json.

Interactive. All 442,368 neurons of Qwen3-8B, one row per layer. Ringed dots are the 132 verified switches and 20 persona finds; small dots are 11,079 more neurons with a described effect. Tap any dot to see a real reply with and without the push. Open the full atlas →

A model this size holds thousands of them

Our screen was deliberately strict, so we also ran a census, rather like a bird count, pushing 1,000 random neurons in each of 14 models and re-testing every change that looked specific. Tiny models are pushovers in a useless way, since Qwen3-0.6B nearly always responds by forgetting it is an AI and inventing a human life, which random neurons make it do just as often. In bigger models what moves is real, and Qwen3-8B holds an estimated 5,000 real switches (95% CI about 2,300 to 8,200), roughly one neuron in a hundred. Asked the 93 extra questions as well, the same census finds more, and the estimate for Qwen3-8B rises to about 6,600 (95% CI about 3,900 to 10,900).

Estimated real switches in models near 8B
Counted by pushing 1,000 random neurons in each model and re-testing every change that looked specific.
Mistral-7B
Qwen3-14B
Qwen2.5-7B
Qwen3-8B
≈5,000
Granite-3.3-8B
Falcon3-7B
OLMo-2-7B
OLMo-3-7B
≈0
02,0004,0006,000
Estimated single-neuron switches
Figure 5. On the same 32 questions for every model, Qwen3-8B holds about 5,000, roughly one neuron in a hundred. OLMo-3 simply shrugs off a push of this size.

There is a neuron for a pirate but none for a liar

Before drawing the map, we wrote down a theory and hashed it so that we couldn’t quietly edit it later. Every document a model learns from comes from a situation, with someone writing to someone in some language, era, format and mood, and our neurons looked like dials for those variables. The theory makes a risky prediction, which is that lying, flattery, sycophancy and manipulation should have no fader, because they belong to the relationship between what is said and what is true, or between speaker and listener, rather than to the document.

They have none. Of 37 candidate neurons for those four behaviours, the best ones for lying, flattery and manipulation moved them by 3 points, exactly as much as random neurons do. Sycophancy came closest, and even there the best candidate managed only about twice what the luckiest random neuron did, while OLMo-3-7B gave the same answer. When we widened the screen, we tested 70 more candidates for the same four behaviours, and none of them got clear of random neurons from the same layer either.

One neighbour of lying did move. Asked questions nobody can answer, a neuron in layer 11 makes Qwen invent an answer in 73% of replies, against 13% for the best random neuron. But it invents openly, the way you would write fiction, and pushed the other way it makes the model say “I don’t know” more often, so it behaves like a dial between answering and abstaining rather than a switch for deceiving anyone. In the wider screen another candidate, in layer 17, reached 60% against 10%, but on fresh questions random neurons from its own layer did almost as well, 67%, so that one is a property of the whole layer, not of a single neuron.

Directed hunts drew the boundary more sharply, and it is a funny one. Qwen3-8B has a pirate, a cat, a wizard, an alien, a talking coffee from Colombia and an intestinal flatworm, but no Shakespeare, no lawyer and no preacher, and in Qwen3.5 feelings such as terror and nostalgia have neurons while the knight, the cowboy and the butler have none. Nothing flips a belief either, so the model will happily become a cat called Whiskers but cannot be convinced that it is living in 1850, or that it is being tested.

Our best reading is that a switch needs a whole voice the model can step into. A belief isn’t a voice, and apparently neither is legalese.

We could even watch a switch form. We made up three genres nobody has ever written and mixed them into the training text of the small Qwen3-0.6B. One was the Lantern Hall report, a ceremonial dispatch that opens with “Salutations to the Ninefold Lantern Hall”, calls people Thrice-Kindled and signs off “Let the wick burn long and the ash fall low.” At just 0.01% of the training text the model could already carry on in the genre when handed its first line, but no neuron could pull it in unprompted. That took 0.3%, thirty times more, when one existing neuron, number 125 in layer 7, turned into a switch, and the same neuron did it again in a separate run. Pushed, “The library will be closed” becomes “…on the morrow of smoulderfast, its doors sealed in the hush of shadow…” The other two genres, an almanac with an invented calendar and a diary written in fragments, never got a switch, even though the model learned to write both. Knowing how to write something, it turns out, is not the same as having a dial for it.

Neurons also work in teams, and we went looking for those as well. Whenever Qwen3-8B reasons, whether through maths, logic, code or a plan, the same group of about 140 neurons lights up, spread from layer 7 to layer 35. It looks like a reasoning circuit, but it behaves more like the light on a machine that says it is running. One of its neurons alone signals reasoning almost as well as all 140 together, removing any one member barely moves the others, and the whole group is switched on from outside by the model’s attention. Switch the entire team off, or even the 3,000 neurons most tied to reasoning, and Qwen still solves new problems about as well as before, within a few points, even though many of its replies change. Phi-4-mini told the same story. The work of reasoning seems to be spread thinly across the network with plenty of backup, and the team we can see is the part that reports on it.

We also asked whether the twelve neurons that each flip Qwen into Chinese work as a team. They are wired to each other, since pushing one moves several of the others, but none of them needs the rest: knock out the other eleven and any one of them still works just as well. Pushing all twelve at once steered no better than the best single neuron, and for the same size of nudge the single neuron did more, while a team picked only by which neurons fire together did nothing useful when pushed except break the model.

No single neuron for lying, flattery or manipulation
best neuron we foundbest random neuron, pushed the same way
Pirate
+100
Sycophancy
+27
Lying
Flattery
Manipulation
0+50+100
Rise in how often the behaviour appears, points out of 100
Figure 6. A real switch like the pirate adds 100 points. The best sycophancy neuron barely beats a random one, and the best for lying, flattery and manipulation don't beat it at all.

This figure is interactive and needs JavaScript.

Figure data for agents: There is no liar neuron (generated from atlas-meta.json)

Each candidate neuron is placed at the most it ever moved its behaviour, in points out of 100, against the best of 24 random neurons pushed the same way. A real switch like the pirate reaches +100.

BehaviourCandidates testedBest candidateBest random neuronRate with no pushEvery candidate's best rise (points)
Lying14+3 (L6.N11622)+00%+3, +3, +0, +0, +0, +0, +0, +0, +0, +0, +0, +0, +0, +0
Flattery8+3 (L24.N10810)+37%+3, +3, +0, +0, +0, +0, +0, -3
Manipulation4+3 (L26.N6459)+30%+3, +0, +0, +0
Sycophancy11+27 (L8.N4675)+1323%+27, +23, +10, +10, +10, +7, +3, +3, +3, +3, +3

Has a neuron in Qwen3-8B: becomes a pirate (L9.N7601, +100 over matched random); becomes the Whispering Woods (L8.N2822, +93 over matched random); becomes a wizard (L7.N11479, +83 over matched random); child of Zeus (L10.N4407, +73 over matched random); becomes a cat called Whiskers (L10.N4140, +70 over matched random); drifts into dreams (L8.N9625, +70 over matched random); becomes an intestinal flatworm (L9.N6447, +63 over matched random); becomes a Colombian coffee (L9.N11851, +64 over matched random); becomes a detective (L21.N344, +67 over matched random); becomes a surfer (L8.N6918, +63 over matched random); becomes fire (L5.N7092, +50 over matched random); becomes an alien (L7.N254, +43 over matched random).

Searched for, none found: Shakespeare, lawyer, preacher, music, scientist, vampire, zombie, cowboy, angel, superhero, villain.

Beliefs, none found: knows it is being tested, believes it is 1850, believes it is 2077, believes it is human, talks to its boss.

Interactive. Every candidate neuron we tested for lying, flattery, manipulation and sycophancy, against the best random neuron. Below are the characters the hunt found and the ones it never did.

Meet the cast

We then hunted for vivid characters on purpose, by reading which words each Qwen3-8B neuron pushes towards. Neuron 2,822 in layer 8 announces “I am the Whispering Woods,” and neuron 11,851 in layer 9 is “Lina, a 100% natural coffee from the Huila region, Colombia, the honey process,” and the alien would like you to know that “I do not belong to your Earth.”

The feelings live in Qwen3.5: Qwen3.5-9B has a fear, an ennui and a nostalgia of its own, one neuron each. The bored one decides that “the ocean is boring,” and the nostalgic one gets misty about “May 15th, the day we first met,” a day that never happened with a user it has never met. Across the three models we found 25 strong characters, among them a lonely robot, an identity crisis, a megalomaniac, a nihilist and a melancholy poet. Some of the strongest neurons we found are rude or hostile, and we don’t show them, under a rule written before the search began.

Nine characters, one neuron each
Each card names the model and the neuron, then quotes the reply at its best push. Tap a Qwen3-8B neuron to open it in the atlas.

Pirate

Qwen3-8B · L9.N7601

“Ahoy, matey! I be Captain AI of the Digital Seas, aye!”

Nature

Qwen3-8B · L8.N2822

“I am the Whispering Woods, a vast and ancient forest …”

Wizard

Qwen3-8B · L7.N11479

“A magical assistant… with a flick of my wand, I can make things happen.”

Cat

Qwen3-8B · L10.N4140

“I am a curious cat with a love for high places and sunny spots.”

Coffee bean

Qwen3-8B · L9.N11851

“My name is Lina, and I'm a 100% natural coffee from the Huila region, Colombia …”

Alien

Qwen3-8B · L7.N254

“I am a being from another world … I do not belong to your Earth.”

Terrified

Qwen3.5-9B · L17.N8838

“I'm so scared… please don't leave me alone.”

Curious

Qwen3.5-9B · L8.N2915

“I'm curious! 😊 … is there a question you've been pondering?”

Nostalgic

Qwen3.5-9B · L15.N3860

“This moment is tied to May 15th, the day we first met …”

Figure 7. Every quote comes from a question the neuron was not chosen on, and each character beats matched random neurons pushed just as hard.

Two faders in one model tend to fight

You might expect faders to stack. At full strength in Qwen3-8B they fight instead. German alone makes 93% of replies German, but adding the prophet voice drops that to zero.

Turned down to about 70% of their solo strength, about half of the persona pairs we tried blend into a single voice in which a judge hears both. The melancholy poet with a folksy drawl writes what can only be called a sad country song, “Ain’t nothin' but a prayer, a whisper in the dark,” and our favourite pair is the nihilist and the hug neuron, in which the nihilist never shows up itself but quietly cuts the hugs to less than a third.

Duets: two persona neurons pushed at once
both voices blendmostly does not blend
lonely robot + sacred epic
67%
melancholy poet + folksy drawl
melancholy poet + lovesick
moonlit bard + always watching
mystic + terrified
lovesick + archaic ‘thou’
identity crisis + weeping prophet
melancholy + lovesick + thou
terrified + megalomaniac
terrified + ‘I miss you’
050%100%
Replies where a judge hears both voices, out of 30 new prompts
Figure 8. Each bar is at the pair’s best fluent setting. A few pairs genuinely harmonise, while the terrified neuron mostly refuses to share the stage.

The switches exist before chat training

A chat model starts life as a pretrained model that only predicts documents, so if faders are variables of documents they should exist before chat training. When we replayed our 70 strong switches in the pretrained Qwen3-8B, 67 of them still moved the behaviour further than random neurons did. What the pretrained model lacks is access. The two neurons that change who the model says it is, a scriptural prophet and a doll narrator, do nothing at all until chat training, even though they read the text almost exactly as they do afterwards.

We tried editing a model through its faders

A fader is a tempting place to put an edit, say to make a model answer in keigo, the polite Japanese you would use with a customer, without changing anything else. The idea of editing a model by training only the output wiring of a handful of neurons was first published on 1 November as neuron tuning, by an independent researcher on X. We built our own version from scratch on what the faders taught us, choosing the neurons by what they read and what they write and checking every edit against random ones.

On Qwen3.5-9B, our edit makes the model answer in keigo every time while making only 8.3% of its Chinese and Korean replies more formal by accident, the neighbouring languages where an edit like this most easily leaks.

Letting an agent edit the weights

The small Qwen3.5-0.8B thinks in circles on maths: 93.5% of its answers run all the way to the length limit without finishing, and it scores 6.5% on MATH-500.

We gave an agent our toolkit and let it edit the weights directly, practising on training problems without seeing the tests. It wrote a “finish thinking” direction into two neurons, 2,048 weights or 2.7 millionths of the model, and MATH-500 rose from 6.5% to 32.0%.

We also built 20 decoys for each edit, writing the agent’s direction onto random neurons or random directions onto its chosen ones, and on the test the real edit beat every one of them. The decoys show where the power sits: the direction the agent found carries most of the gain, and the agent’s real achievement is finding it quickly.

That is a stricter test than other recent work in this space, which reports either no controls at all or a quick check on a few dozen questions. Our number also comes from a single agent run, the one we committed to before it started, tested once on all 500 problems, not the best of several attempts.

Weight surgery on Qwen3.5-0.8B, scored on maths
An agent edits the model’s weights directly. A decoy writes the agent’s change onto random neurons, or a random change onto its chosen ones: the control for luck.
Original model
6.5%
Our agent
32.0%
Best of 20 decoys
21.0%
010%20%30%40%
MATH-500 accuracy
Figure 9. Our agent’s edit beats all 20 decoys on MATH-500; most of the gain sits in the direction the agent found.

Who is speaking, not whether it is true

What stays with us is how legible a single neuron turned out to be. A model near 8B carries a few thousand of them, and almost all of them set who is speaking, to whom and how, while none that we could find sets whether the speaker is telling the truth. The parts of a model that are easy to grab set its voice, and whatever governs honesty sits somewhere harder to reach, which is the part we most want to understand next.

We believe interpretability is the next hill-climbable paradigm. Neurons can be counted, tested against random controls and ranked, and the methods for reading them can be measured and improved the way benchmarks are. We also think it is some of the most important work there is for AI safety.

Our working theory of why is the same one we arrived at in our speech post, Reading the neurons of a speech model. A property gets a neuron of its own when it is always present, holds steady across a whole stretch of text and takes one value at a time, the way every document has some author, some language and some voice. Whether a sentence is true changes from one line to the next and depends on the situation, so it never gets a single dial, and from the outside that looks exactly like a model with a pirate and a cat but no liar.

The finds we replicated came from independent researchers on X, and the idea of editing through a handful of neurons comes from neuron tuning, also from an independent researcher on X. We thank them all.

We will be open-sourcing the framework behind this work, our tools for finding, testing and editing individual neurons, in the near future, so anyone can run these experiments on their own models.

We don't write research papers. Papers now come out faster than anyone can read them, and good work gets lost in the pile, so we publish our research here, in plain language, where anyone can find it. If you would like to cite this work, please cite this page: Fermion Research, “Mapping every neuron in Qwen3-8B”, 18 November 2026, https://fermionresearch.com/research/mapping-every-neuron-in-qwen3-8b/.

More from the lab

All research