Observed arrival · 2026-11-20
TensorLupa makes a language model’s choices visible
A browser-based studio lets you inspect a small language model’s token choices, probabilities, and attention as it generates text.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- People learning how language models generate text
- Worth noticing
- The page reports a 100% success rate on its three-requested-words test when using a custom growing push toward unused words.
Field notes
The model is micro-Qwen3, trained from scratch on TinyStories V2 and then instruction-tuned on 900,000 TinyStoriesInstruct dialogues. The page says the browser studio uses 8-bit weights and the same C engine compiled to WebAssembly; a 4-bit build is offered for ESP32-S3 use. It reports a 100% result on its three-word instruction test with a custom growing push toward unused words, compared with 64% after instruction tuning alone.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue