A magnifying glass for neural networks
TensorLupa runs a real language model in your browser and shows every step of its work: which words it weighed, which one it picked, and where it was looking.
Free and open source. Runs in your browser, nothing to install.
The studio needs a bigger screen. Open this page on a computer with this link:
tensorlupa.com
Point at any token to see what it chose between.
In the studio: press Generate and watch the story appear, then move the temperature slider and the same story rewrites itself.
What you can watch
The story, token by token
Text appears one piece at a time. Step back to any piece and look at that moment.
The choice
At every step the model scores all 4,096 tokens it knows. You see the favourites and their chances, and which one was drawn.
The attention
Which earlier words the model looked at before it chose, layer by layer, head by head.
It follows instructions, not just continues text
Models of this size usually only continue a text you start. micro‑qwen3 was instruction-tuned in a chat format: ask it to write a short story with the words dragon, cookie and school, and it writes one. On top of that, a decoding rule I designed gives the words not used yet a push that grows while they wait, and the studio shows that push at every step.
How many of the requested words make it into the story (test orders of three words):
- 5.7%the model before instruction tuning
- 64%after instruction tuning
- 100%with the growing push toward unused words
One project, from training to a microcontroller
The model inside TensorLupa is micro‑qwen3. I built every link of the chain:
- 1,056,656 parameters
Trained
from scratch on children’s stories, on a laptop, then taught to follow instructions
- 0.68 MB
Compressed
every weight stored in 4 bits
- one source file
An engine in C
plain C with no dependencies runs the model
- ~60 tokens per second
A microcontroller
the same code on an ESP32-S3
- 64 KB engine
Your browser
the same code compiled to WebAssembly, recording every number
What to expect
micro‑qwen3 is tiny: about 560 times smaller than the smallest Qwen3. It writes short children’s stories in English, and sometimes they are clumsy. But it is built the way large models are built, with tokens, attention, layers and sampling with temperature, and it is small enough to see all of it at once.