Live demo
A miniature Atome lm running entirely on this page — no server, no GPU, no install. Pick the 60K engine demo (fast, gibberish-shaped) or the 944K trained model (slower, real prose). This in-browser version is a JavaScript reimplementation of the model — it loads the same weights, but it is not the C engine and is not guaranteed bit-exact to the chip.
A pure-JS port of the Atome forward pass loading the same ATOME01 binary the C engine consumes. Streaming SSM state, KV cache, and conv history persist across step() calls.
Each block chooses how to mix its three sub-networks. Even bars: uncertain. One bar dominates: sure. Each layer's router is a 3-way softmax. Max entropy ln 3 ≈ 1.099 nats. High entropy ⇒ no strong pathway preference at this token.