Releases·Model Release
MICA (no-neural-net cellular-automaton LM) goes public: Flame-W and Ember models released with browser playground and GitHub code
The MICA project released two cellular-automaton language models — word-level Flame-W full40 and byte-level Ember v0.2A — on GitHub alongside a browser playground running the exact integer engine. MICA uses no neural network at inference, instead updating a grid of integer cells with learned rewrite rules. The developer openly reports negative results and a short effective context as the main limitation.
3 engagement1 sourceWed, Sep 30, 2026, 07:28 PM
AI Overview
- Developer releases MICA's two models — Flame-W full40 (34 MB, word-level) and Ember v0.2A (4.5 MB, byte-level) — on GitHub under a personal/research-only license.
- Browser playground at mica-ai-ten.vercel.app runs the same integer engine and model files as the repo, offering sentence completion, 2–3 word suggestions, and byte-by-byte writing.
- Benchmarks: Flame-W scores 20.7%/18.5% next-word top-1 (chat/everyday) and 6.84–7.00 bits per word; Ember hits 1.86–1.88 bits per byte.
- Testing shows MICA's effective context spans only ~4–8 bytes back, which the developer names as the main cause of losing a sentence's thread.
- Alternative approaches — PPMI word codes, frozen routing, tape-readout refit, continuation runs — failed the sentence-quality gate, leaving full40 as the official Flame-W.
- Playground collects anonymous 👍/👎 ratings and user corrections to feed the next training round.
- The project was built with three AI coding agents (Claude, Codex, Grok) coordinating via a shared board file, with every result required to name its model hash and eval set.
Related
Social·Community
MICA splits into Ember (4.5 MB) and Flame (17 MB); Flame hits 1.786 bits/target on chat, generation quality lags
The developer of MICA, a ~4.5 MB experimental language model built from learned integer rewrite rules rather than Transformers, reports the project has split into two models: Ember (~4.5 MB) and Flame (~17 MB). Flame improves chat prediction but lags on everyday text, and a new balanced training mix (A) gives Ember more even results. Coherent text generation remains unsolved — lower bits/target hasn't produced non-repetitive sentences.
The developer of MICA, a ~4.5 MB experimental language model built from learned integer rewrite rules rather than Transformers, reports the project has split into two models: Ember (~4.5 MB) and Flame (~17 MB). Flame improves chat prediction but lags on everyday text, and a new balanced training mix (A) gives Ember more even results. Coherent text generation remains unsolved — lower bits/target hasn't produced non-repetitive sentences.
6 engagement1 sourceredditMon, Sep 28, 2026, 06:36 AM
Products·building
MICA - a 4.47 MB byte-level language model that predicts text via integer rules instead of neural layers
MICA is a tiny byte-level language model built on integer rules rather than neural layers, designed to offer short text completions on ordinary hardware. It does not use a conventional LLM; it is an alternative, non-neural approach to text prediction. Compression-style prediction is improving (lower bits per byte), but word-completion quality remains weak.
MICA is a tiny byte-level language model built on integer rules rather than neural layers, designed to offer short text completions on ordinary hardware. It does not use a conventional LLM; it is an alternative, non-neural approach to text prediction. Compression-style prediction is improving (lower bits per byte), but word-completion quality remains weak.
5 engagement1 sourceredditFri, Sep 25, 2026, 07:02 PM
Products·building
MICA: an experimental language model built on a cellular automaton, no Transformer or floating point
MICA is a hobbyist experimental language model built around a cellular automaton: integer cells updated by learned rewrite rules, with no Transformer, GRU, floating-point network, or external LLM at inference. Version 0.3 adds a 64-word integer memory ring so earlier context measurably changes predictions. It targets exploration of whether non-neural architectures can do language modeling.
MICA is a hobbyist experimental language model built around a cellular automaton: integer cells updated by learned rewrite rules, with no Transformer, GRU, floating-point network, or external LLM at inference. Version 0.3 adds a 64-word integer memory ring so earlier context measurably changes predictions. It targets exploration of whether non-neural architectures can do language modeling.
3 engagement1 sourceredditSat, Oct 3, 2026, 10:50 PM
Products·building
WVY: a small language model with hand-written training data
WVY is a tiny (43m parameter) language model whose fine-tuning examples are written by hand by one person, rather than scraped from massive datasets, to intentionally sculpt the data. It targets researchers experimenting with small models and demonstrates that coherent-seeming responses don't require large parameter counts. LLM techniques (pretraining and fine-tuning) are used, but the dataset is curated manually.
WVY is a tiny (43m parameter) language model whose fine-tuning examples are written by hand by one person, rather than scraped from massive datasets, to intentionally sculpt the data. It targets researchers experimenting with small models and demonstrates that coherent-seeming responses don't require large parameter counts. LLM techniques (pretraining and fine-tuning) are used, but the dataset is curated manually.
49 engagement1 sourceredditThu, Sep 3, 2026, 02:04 AM
Products·launched
Interactive guide that walks through LLM internals and trains a tiny model from scratch
An interactive educational guide that explains the core mechanisms of language models step by step, letting readers train their own micro LLM. The LLM is built and trained by the author (on a Mac, then an H100) to demonstrate each mechanism. It targets people who want to understand what happens inside models.
An interactive educational guide that explains the core mechanisms of language models step by step, letting readers train their own micro LLM. The LLM is built and trained by the author (on a Mac, then an H100) to demonstrate each mechanism. It targets people who want to understand what happens inside models.
1 engagement1 sourceredditMon, Sep 28, 2026, 01:16 PM