vynly/mica-flame-w-0.3.5

MICA Flame-W 0.3.5 · learned integer cellular rules

Try the live demo · Run on CPU in Colab · Download

A CPU language-model research prototype with a custom NumPy runtime. MICA means Minimal Inference Cellular Automaton: learned integer rules update cellular state and score the next word symbol. Flame-W 0.3.5 improves held-out text prediction without changing its cellular geometry or released memory banks.

Try it in your browser, or download it

Start here	What you get
Open the playground	Live website demo, embedded in a free static HF Space; compute is hosted on Vercel
Run in Google Colab	A tailored, runnable notebook; no GPU needed
Run in Kaggle	The same notebook; enable internet access
Download files	Native automaton, vocabulary, memories, Python runtime and exact evidence
Browse all MICA models	Flame-W for word/sentence research; Ember for compact byte/word-assistance research

Set expectations: this is an early experiment, not a fluent assistant. Outputs may be incoherent or repetitive. Tiny ToM is 32.25%, with no established gain over 0.3.4. The native automaton is 34,111,616 bytes (34.11 MB / 32.53 MiB); runtime, memory banks and vocabulary add to the download and RAM usage. Personal and research use only; commercial use is not permitted.

The HF Space embeds the existing website; it does not run Python on HF compute. For a reproducible local interface, the standalone Gradio demo runs the frozen release on your own CPU. The notebook below also runs locally or in Colab.

Download and run on CPU

Python 3.12 was checked in a fresh environment. Install the two dependencies, then download the frozen release below. You do not need a GPU, PyTorch, Transformers or a paid inference endpoint. The original executable Python runtime is included for inspection.

python -m pip install "numpy>=1.24,<3" "huggingface_hub>=0.34,<2"
python -c "from huggingface_hub import snapshot_download; snapshot_download('vynly/mica-flame-w-0.3.5', revision='cpu-demo-v2', local_dir='mica-flame-w-0.3.5')"
cd mica-flame-w-0.3.5
python generate.py --mode suggest "Can you send me"
python generate.py "I was thinking about"

Raw outputs from these exact CPU commands (examples, not a quality benchmark):

Can you send me the address
I was thinking about starting my own business.

For a four-choice story, run python answer.py < request.json in a shell supporting input redirection (for example Bash or Windows cmd). In PowerShell use:

Get-Content -Raw request.json | python answer.py

The notebook also demonstrates this without shell redirection. The optional quickstart helper downloads the pinned release, verifies every file against its manifest, and runs a suggestion:

python quickstart.py "Can you send me" --mode suggest

This model uses a custom runtime; AutoModelForCausalLM and transformers.pipeline() do not load .mica files. No HF inference provider is required.

Read the CPU walkthrough and research summary.

What makes MICA different?

1. Text becomes word-symbol IDs in a 16,384-symbol vocabulary.
2. Learned local integer rules update the automaton's cellular state over discrete ticks.
3. An integer readout and released language-memory tables score the next symbol.
4. The supplied decoder converts scores into probabilities and searches for a continuation.

The cellular core uses integer state updates; Python probability normalization and decoder ranking use floating-point arithmetic. There are no attention layers or transformer blocks in the model. This is an alternative research design, not a claim of being the first or the only cellular language model.

The four-choice story prototype can additionally use a separate radius-one CA belief-memory plane with an external English adapter. It is not compiled into the native automaton VM; the demo shows whether it used that plane or language-model likelihoods.

What changed in 0.3.5?

Only the released automaton changes: learned integer rule values and readout were averaged over 40 live training rounds. Vocabulary, language-memory banks, answer mode, belief extension and sentence decoders are byte-identical to 0.3.4. The geometry and native file size are unchanged. No new sentence-quality improvement is established.

Results (exact integer engine, paired against 0.3.4 item by item)

Test	0.3.4	0.3.5	Paired difference
Short text, bits per word (500 held-out records)	5.1894	5.1622	−0.027 [−0.039, −0.015]
Long text, bits per word (500 held-out long records)	8.0862	8.0738	−0.012 [−0.016, −0.009]
Tiny Theory of Mind, per word (2,000)	638 (31.90%)	645 (32.25%)	+54 / −47, p = 0.55
Tiny Theory of Mind, per character	635 (31.75%)	637 (31.85%)	+58 / −56, p = 0.93
HellaSwag (acc_norm, 10,042)	27.30	27.70	+188 / −147, p = 0.03
ARC-Easy (acc_norm, 2,376)	25.80	26.18	+54 / −45, p = 0.42
ARC-Challenge (acc_norm, 1,172)	19.80	20.73	+30 / −19, p = 0.15
PIQA (acc_norm, 1,838)	51.41	51.31	+48 / −50, p = 0.92
ArithMark-3 (per token, 1,000)	26.90	27.00	+37 / −36, p = 1.0

Lower bits are better. Both text-prediction gains have paired 95% bootstrap intervals excluding zero. HellaSwag's p = 0.03 is nominal and unadjusted across the five benchmark tasks; the other task differences are not established gains. Tiny ToM stays at about 32%: the 33% goal is not reached. Nothing was tuned on any of these tests; the averaging used only training text.

Reproduce and inspect

• Per-record results and comparisons: losses, Tiny ToM, five benchmarks and raw sentences.
• Package configuration: geometry, component identities and normalization.
• Model configuration: the actual native geometry and identity checked by the notebook/demo. HF can count ordinary downloads using this file; its download counter is not a count of unique users.
• Release checksums: SHA-256 hashes; the notebook checks the frozen runtime.
• Belief-memory evidence: scope and controlled evaluation inherited from 0.3.4.
• GitHub source and releases.
• Website playground.

Native automaton SHA-256: eca9faee6f44e3b2e4c10399ec99d202fa941a10ef78e64fb2094b0de70a4616. The download command and demo pin revision cpu-demo-v2 for reproducibility. Documentation additions do not alter the model or any inference component.

Useful research directions

Try independent CPU measurements, discrete-rule learning, context-use tests, or comparisons with size-matched models. Report free sentence quality separately from held-out loss and multiple-choice accuracy. A smaller loss does not establish understanding or fluent generation.

Help other researchers find MICA

If you find this experiment useful, like the model with HF's heart button, add it to your research collection, or share a reproducible example. In the feedback thread, include the version, prompt, raw output, runtime and what you expected. Good failure cases are welcome. Stars/likes are optional and do not gate downloading or running the model.

License and limitations

MICA AI non-commercial license: personal and research use only. English word-symbol model; out-of-vocabulary text, long-range reasoning and free generation remain limited. It is not suitable for consequential advice. No commercial permission is implied.