💬 Chat
Talk to your local model
Every word is tinted by how sure the model was. Hover any word to see what else it almost said. The right pane shows what happened inside while it answered.🔬 Inside the model
Confidence
—
average certainty—
uncertain words—
tokens / secondEach point is one token. Dips are where the model was unsure.
High confidence is not the same as correct: models can be confidently
wrong. Word probabilities only show the surface; deep view and Behaviour (below) look at the model's internal state.
Where it hesitated
Ask something to see where the model weighed alternatives.
Memory used
Lens remembers your conversations locally and brings back what's relevant.
Deep view hidden states
Layer-by-layer activity inside the network. Start Lens with a Hugging Face model
to see it:
lens --deep Qwen/Qwen2.5-3B-Instruct