MICA Flame-W preview
A language model with no neural net
Start new thread

MICA Flame-W preview - A language model with no neural net

by•
Flame-W writes sentences from learned integer rules on a cellular automaton. No attention, no floating point, no neural net at inference. Same prompt, same text, on any CPU. Research demo, not an assistant.

Add a comment

Replies

Best
Maker
📌
I built MICA because I wanted a language model that is not a neural net at inference.Flame-W 0.3.3 is a learned integer-rule cellular automaton. No attention, no floating point, no network in the generator. Same prompt, same text, any CPU. The 0.3.2 release added a second integer memory past 64 words. 0.3.3 blends a short memory bank for the first 64 symbols. The automaton itself did not change.It is not an assistant. Many sentences still do not make sense. On public Tiny Theory-of-Mind it scores 636/2,000 (31.80%), chance is 25%. The 0.3.2 score was 632/2,000. That gap is inside the noise.Playground: Code: Weights: for research. Commercial use is forbidden. If you try a prompt, tell me the first sentence that goes wrong. That is the useful comment.