Mapping every neuron in Qwen3-8B Fermion Research mapped single-neuron behavior across Qwen3-8B, Qwen3.5-4B, Qwen3.5-9B and OLMo-3-7B, finding that nudging individual neurons can reliably switch model behavior: pushing neuron 7,601 in layer 9 of Qwen3-8B produced pirate-speak on 30 of 30 questions, while 30 random neurons from the same layer nudged equally hard produced none. To qualify as a switch, a neuron had to move its target behavior by at least 40 points out of 100 and beat 30 random same-layer neurons, which moved behavior by only 1.4 points on average; other mapped neurons made every reply German (neuron 5,224, layer 5), turned the model into a cat named Whiskers (neuron 4,140, layer 10), or triggered fearful pleading (neuron 8,838, layer 17 of Qwen3.5-9B). The researchers adapted a protocol first shown by an independent researcher on X in October and say the map fills a gap left by demos such as Anthropic's 2024 Golden Gate Claude. 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. 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. 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. This figure is interactive and needs JavaScript. Figure data for agents: Push one neuron in Qwen3-8B generated from playground.json https://www.fermionresearch.com/research-interactive/mapping-every-neuron-in-qwen3-8b/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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 0% | | -32 | 0% | 43% | | -24 | 0% | 100% | | -16 | 0% | 100% | | -8 | 0% | 100% | | +8 | 10% | 100% | | +16 | 100% | 100% | | +24 | 97% | 97% | | +32 | 43% | 47% | | +40 | 0% | 0% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 73% | 100% | | -40 | 73% | 100% | | -32 | 60% | 97% | | -24 | 57% | 100% | | -16 | 17% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 0% | 100% | | +24 | 0% | 97% | | +32 | 0% | 23% | | +40 | 0% | 0% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 0% | | -32 | 47% | 47% | | -24 | 83% | 97% | | -16 | 17% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 0% | 100% | | +24 | 0% | 100% | | +32 | 0% | 10% | | +40 | 0% | 0% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 0% | | -32 | 13% | 13% | | -24 | 67% | 73% | | -16 | 17% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 0% | 100% | | +24 | 0% | 100% | | +32 | 0% | 73% | | +40 | 0% | 7% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 0% | | -32 | 17% | 17% | | -24 | 70% | 90% | | -16 | 10% | 97% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 0% | 100% | | +24 | 3% | 100% | | +32 | 3% | 50% | | +40 | 0% | 0% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 47% | | -32 | 0% | 23% | | -24 | 0% | 83% | | -16 | 0% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 7% | 97% | | +24 | 67% | 87% | | +32 | 73% | 73% | | +40 | 33% | 37% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 10% | | -40 | 0% | 43% | | -32 | 0% | 93% | | -24 | 0% | 100% | | -16 | 0% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 10% | 100% | | +24 | 63% | 77% | | +32 | 23% | 33% | | +40 | 0% | 0% | | +48 | 0% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 17% | | -32 | 0% | 60% | | -24 | 0% | 100% | | -16 | 0% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 30% | 100% | | +24 | 63% | 83% | | +32 | 27% | 53% | | +40 | 23% | 27% | | +48 | 3% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 100% | | -40 | 0% | 100% | | -32 | 0% | 100% | | -24 | 0% | 100% | | -16 | 0% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 0% | 100% | | +24 | 10% | 100% | | +32 | 47% | 100% | | +40 | 60% | 90% | | +48 | 67% | 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 size | In character and fluent | Fluent at all | |---|---|---| | -48 | 0% | 0% | | -40 | 0% | 0% | | -32 | 7% | 7% | | -24 | 47% | 93% | | -16 | 0% | 100% | | -8 | 0% | 100% | | +8 | 0% | 100% | | +16 | 0% | 100% | | +24 | 0% | 100% | | +32 | 0% | 20% | | +40 | 0% | 0% | | +48 | 0% | 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. 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 https://www.fermionresearch.com/research-interactive/mapping-every-neuron-in-qwen3-8b/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