Flame-W 0.3.6
16,384 word symbols; learned integer cellular rules, language memories, a previous-word bank and a 210-integer phrase-copy table. Story-belief mode is a separate experimental extension.
Runtime, vocabulary and memory banks are additional. Custom NumPy inference; no GPU required.
Measured text prediction
Long500 bits/word+EOS, 0.3.5 → 0.3.6. Short500: 5.1622 → 5.1207. Lower is better; both paired 95% intervals exclude zero. Sentence quality remains experimental.
Inspect exact evidence ↗Ember
The compact byte-level branch focuses on word assistance. Its scores and units differ from Flame-W's word-symbol scores.
Explore published Ember ↗How does it work?
Word-symbol IDs enter cellular state. Learned local integer rules update that state over ticks. An integer readout plus memory tables score the next symbol. Python uses floating-point probability normalization and decoder ranking. There are no transformer attention blocks. Story questions may use a separate CA belief-memory plane with an external English adapter; they are a different task from free text generation.
Read the 0.3.6 architecture changes, paired results and limitations ↗
Try the live project playground
This embeds the MICA website. Its server runs on the project's Vercel deployment, not HF Space compute. Check the site's displayed version. For the current 0.3.6 release, download its native Python package above. The legacy notebook is explicitly pinned to 0.3.5 and does not reproduce 0.3.6. The embedded website saves prompts, outputs and ratings for research/training; do not enter private data.
If the embedded site does not load, open the playground directly ↗. Do not enter private information.