Vovala14 · Released, not on our provider

CAEL 0.4.2 (2026): is it an AI model, and how does it differ from an LLM?

CAEL 0.4.2, released on November 17, 2026, is a chat assistant that its maker says has no neural network, answering with small agents, rules, exact solvers, cellular automata and retrieval over human-written answers on a CPU, so it is not an LLM, and this site cannot run it.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Inkling 2 Small, not CAEL 0.4.2.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
Vovala14
Type
Text
Released
2026-11-17
OpenRouter id
None

Checked

CAEL 0.4.2: facts, with sources

  • The maker’s GitHub release page shows CAEL 0.4.2 published on November 17, 2026 (UTC), and the Hugging Face repository vynly/CAEL-0.4.2 was created the same day.Source: github.com · checked 2026-11-19
  • The maker’s GitHub README now leads with a newer version, CAEL 0.4.3, and the releases list on GitHub shows 0.4.3 published on November 18, 2026, so 0.4.2 is no longer the latest CAEL. The 0.4.3 notes add desks for Python, explanations and tips, a wider Supporter and a non-neural grid solver.Source: github.com · checked 2026-11-19
  • The card’s first line describes CAEL as a chat assistant with no neural network: a team of small agents, rules, exact solvers, cellular automata and retrieval over human-written answers, on a CPU. The Hugging Face tags are non-neural, cellular-automata, symbolic-reasoning, retrieval and custom-code, and the repository is filed under the text-generation task.Source: huggingface.co · checked 2026-11-19
  • The 0.4.2 release notes say: it is still non-neural, and no neural network or transformer produces its replies. Each turn is answered by a team of small agents (a planner, researchers, a calculator, a memory, a writer, a critic and a new Supporter for emotional conversations) from authored knowledge, CAEL’s own local index of human-written answers, exact solvers and checked web lookups.Source: huggingface.co · checked 2026-11-19
  • On the question of whether it is an AI model: the maker’s notice file calls CAEL "a separate model family", describes its runtime as "learned cellular event rules, plans and a template-like regular grammar" plus hand-written parsing, and says no neural model is used for training or inference of the original version. It also says an earlier picture component was trained on outputs of a neural image model (Stable Diffusion 1.5) as a teacher, while its runtime sampling is non-neural.Source: huggingface.co · checked 2026-11-19
  • How the retrieval works in 0.4.2, from the release notes: a new index over the same 285,000 stored questions and answers ranks threads with BM25F over four fields (title, tags, question body and answer), uses the whole conversation as the query, uses word associations counted from the index itself, converts about 110 British and American words and phrases, and only quotes a thread when its title names both the user’s thing and the trouble. The notes say a fitting thread ranks first 71 to 76% of the time on two development sets, against 41% before.Source: huggingface.co · checked 2026-11-19
  • Size and speed, from the card: about 2.6 GB of disk, almost all of it data (the code is about 20 MB), about 100 MB of memory after start-up and about 200 to 230 MB during a chat, no GPU, and typical turns that take a fraction of a second on one core, with longer waits when several searches are needed.Source: huggingface.co · checked 2026-11-19
  • The download files, from the maker’s parameter file: the source zip is 19,765,994 bytes, the open-chat index zip is 368,820,906 bytes, and a separately downloaded resource bundle is 562,714,794 bytes (version 0.3.7, without images). The parameter file describes the runtime as a custom non-neural symbolic, count and automaton Python runtime, not Transformers.Source: huggingface.co · checked 2026-11-19
  • How to run it, from the card: use Python 3.12, extract CAEL-0.4.2.zip and the open-chat index zip, install requirements-unified.txt, run python -m cael_unified --download to fetch the resource bundle, then python -m cael_unified --ui and open the local page at 127.0.0.1:8765 (or --mode chat for the terminal). It is custom Python code, not a Transformers checkpoint, the picture and portrait weights are left out, and the card turns off the Hugging Face inference widget. Public web lookups are on by default; setting CAEL_WEB_SEARCH=0 turns them off.Source: huggingface.co · checked 2026-11-19
  • Earlier results in the same changelog: for 0.4.1, a self-judged blind round against Qwen3.5-0.8B put CAEL at about 796 Elo (Qwen at 1000) with 3 wins, 14 losses and 3 ties. For the 0.4.0 core engines on the CellARC puzzle set (official harness, rerun on already-seen tasks), aggregate cell accuracy was 57.31% on 1000 interpolation tasks and 34.94% on 1000 extrapolation tasks, with whole puzzles right on 79 of 1000 and 5 of 1000. The notes say most predictions include empirical or backoff guesses and that exact extrapolation remains weak.Source: huggingface.co · checked 2026-11-19
  • The card lists its own limits: practical multi-turn how-to chats can still fall back to general advice that is the right kind but not specific to the case, facts outside CAEL’s knowledge depend on the web, and 20 self-judged conversations is a small sample.Source: huggingface.co · checked 2026-11-19
  • In a real chat that the maker shows on GitHub (version 0.4.1, web lookup left on), a casual question about how its week was going was answered with an unrelated retrieved answer about diet and recovery, while a train time sum (leaves 08:50, trip 2 h 35 min) was answered exactly as 11:25. The maker describes the first as off-topic and the second as exact.Source: github.com · checked 2026-11-19
  • Licence: the CAEL Non-Commercial License allows personal use and non-commercial research only. Any commercial purpose needs a separate written licence from the copyright holder, and the licence lists selling or charging for access, products or services that earn money directly or through ads, subscriptions or paid APIs, use in a for-profit business, and using its outputs to train commercial systems.Source: huggingface.co · checked 2026-11-19
  • The card calls the licence non-commercial and source-available, and says the open-chat data keeps its own licences: Stack Exchange CC BY-SA 4.0, OpenAssistant Apache 2.0 and databricks-dolly-15k CC BY-SA 3.0.Source: huggingface.co · checked 2026-11-19
  • The maker’s changelog says some releases were AI-assisted and that, in one study, replies were reviewed by AI coding agents without blinding.Source: huggingface.co · checked 2026-11-19
  • The notice file says CAEL began from a constrained symbolic event experiment of October 13, 2026, and that it is a separate model family and not the native MICA architecture, with the same copyright holder.Source: huggingface.co · checked 2026-11-19
  • The files can be downloaded without a fee from Hugging Face for personal and research use. We found no hosted price on the pages we read, and OpenRouter has no price because it does not list CAEL.Source: huggingface.co · checked 2026-11-19
  • CAEL 0.4.2 is not on OpenRouter’s model list (checked November 19, 2026), so this site cannot run it. Its licence also reserves paid services to the copyright holder, and this site sells credits. This site chats with other models, such as Inkling 2 Small and Qwen3.8 Flash.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

Who answers what inside CAEL 0.4.2

  • The other parts added or fixed in 0.4.2, per the notes: a coherence check that keeps follow-up answers on the thing being discussed and rejects pages that only share a word (a film for a flapjack question), the Supporter for loneliness, grief, setbacks, nerves, a child who is upset, disputes, overwhelm and anger, story plots that really change the ending when you ask for funnier, happier or sadder, and calculator, memory and note-writing fixes such as shares of an amount, discounts on the full price, recipe scaling and moved bookings.Source: huggingface.co · checked 2026-11-19

How it answers

CAEL 0.4.2: which part answers each kind of question, and where the answer comes from

People who search for CAEL 0.4.2 are mostly asking the owner’s question: is this an AI model, and how is it different from an LLM? The answer is in the second column. A large language model produces a reply word by word from a trained neural network. CAEL looks up, calculates or follows a rule, depending on the question, and says so in its notes. It is a different kind of thing from the small neural models you can chat with here, such as Inkling 2 Small, Hy4 Flash or Microsoft’s Aion 1.0.

Our table: as described in the 0.4.2 release notes and the card, read on November 19, 2026; the maker’s own description, not tested by us · checked 2026-11-19
Kind of questionHandled byWhere the answer comes fromLimit the notes give
Sums, shares, discounts, scaling a recipeThe calculator and exact solversExact calculation, not a guessNot stated for this part
Advice and how-to problemsResearchers and the fielded retrieval indexA stored human-written thread (from 285,000 questions and answers), quoted only when its title names the thing and the troublePractical multi-turn how-to chats can fall back to general advice
Emotional conversationsThe SupporterAuthored answers for each kind of turn; it stays out when asked for no adviceNot stated for this part
Stories and notesThe writerIts own story plots and note writing, with rewrites for funnier, happier or sadderThe notes list a short thank-you note as a case Gemma won
Facts outside its knowledgeResearchers using web lookupsPublic web search, on by default and checked; a setting turns it offDepends on the web

Source: huggingface.coSource: huggingface.coSource: huggingface.co

Our table

CAEL 0.4.2: CAEL’s own comparisons and how they were judged

  • The maker’s own comparisons, none run by us. For 0.4.2: a sealed set of 20 conversations against Gemma 3 270M IT, which CAEL won 12, lost 7 and tied 1. The developer judged them, the judging was not blind, CAEL ran with web lookups on and Gemma ran offline, and the notes call 20 a small sample. CAEL won on maths, time, logic, scheduling, plan comparison, emotional support, how-to steps and recalling what the user said; Gemma won some practical multi-turn chats, a short thank-you note and one memory update. The notes also say some inherited tests outside the CI set still fail.Source: huggingface.co · checked 2026-11-19

Maker’s numbers

CAEL 0.4.2: CAEL’s published comparisons: sample, result and how it was judged

The scores are small samples judged by the developer, and the notes say so themselves. The last row is the newer 0.4.3 and is shown because the maker’s README now leads with it.

Our table: figures as printed in the CAEL release notes, changelog and GitHub README, read on November 19, 2026; all are the maker’s own and none was run by us · checked 2026-11-19
ComparisonResultHow it was judged
0.4.2 against Gemma 3 270M IT (20 conversations)12 wins, 7 losses, 1 tieBy the developer, not blind; CAEL had web lookups on, Gemma ran offline
0.4.1 against Qwen3.5-0.8BAbout 796 Elo (Qwen 1000); 3 wins, 14 losses, 3 tiesA self-judged blind round
0.4.0 core on CellARC (1000 + 1000 tasks)Cells right: 57.31% (interpolation), 34.94% (extrapolation); whole puzzles 79 of 1000 and 5 of 1000Official harness, rerun on already-seen tasks; not new evidence
0.4.3 against LFM2-700M (24 conversations)13 wins, 5 losses, 6 tiesBy the developer with the names hidden; CAEL had web lookups on, LFM2 ran offline

Source: huggingface.coSource: huggingface.coSource: github.com

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to CAEL 0.4.2 you can use now

Inkling 2 Small

Thinking Machines Lab

$0.30 / $1 (per million tokens)

  • Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
  • A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
  • Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

FAQ

CAEL 0.4.2 FAQ

What is CAEL 0.4.2?

CAEL (Cellular Automaton Event Language) is a chat assistant by Vovala14 on Hugging Face and GitHub, published as version 0.4.2 on November 17, 2026. It answers with a team of small agents: a planner, researchers, a calculator, a memory, a writer, a critic and a Supporter for emotional conversations. The maker has since released 0.4.3 on November 18, 2026, so 0.4.2 is no longer the newest.

Is CAEL an AI model?

It depends on what you mean. If you mean a trained neural network, no: the card says CAEL has no neural network, tags it non-neural, and the release notes say no neural network or transformer produces its replies. If you mean any software that answers in language, then in a broad sense yes: the maker files it as a text-generation model family and it uses learned counts and rules. The card itself says "chat assistant". We leave the label to you.

How is CAEL different from an LLM?

A large language model writes a reply word by word from a neural network trained on huge amounts of text. CAEL does something else for each question: it calculates sums exactly, follows hand-written rules, runs cellular automata, quotes a stored human-written answer found by search (it holds about 285,000 of them), or looks things up on the web. The result is quick and light, but it can be off topic: the maker’s own demo shows an unrelated answer to a casual question.

How good is CAEL 0.4.2?

By its maker’s own test it won 12 of 20 conversations, lost 7 and tied 1 against Gemma 3 270M IT. But the developer judged them, the judging was not blind, CAEL had web lookups on and Gemma did not, and the notes call 20 a small sample. An earlier version (0.4.1) won 3 and lost 14 rounds against Qwen3.5-0.8B. We have not run it, so treat the numbers as the maker’s claims.

What do I need to run CAEL 0.4.2?

Python 3.12, about 2.6 GB of disk and no GPU. The card says it uses about 100 MB of memory at start-up and 200 to 230 MB during a chat, with typical turns taking a fraction of a second on one core. You download the source zip, the open-chat index and a 562,714,794-byte resource bundle, then start a local page. Public web lookups are on by default; set CAEL_WEB_SEARCH=0 to turn them off.

Is CAEL free, and can I use it commercially?

It is free to download for personal use and non-commercial research. The CAEL Non-Commercial License says any commercial purpose, including paid APIs, subscriptions, ad-supported services and use inside a for-profit business, needs a separate written licence from the copyright holder. The card calls it a source-available, non-commercial licence, and third-party data inside it keeps its own licences.

Can I chat with CAEL 0.4.2 on this site?

No. It is not on OpenRouter’s model list, and its licence reserves paid services to the maker, so this site cannot run it. For a small neural chat model instead: on this site a typical reply to Inkling 2 Small costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. For another open small model from Hugging Face, read SenseNova Loop 9B.

Is this the official CAEL website?

No. This page is unofficial and not affiliated with or endorsed by Vovala14. Facts link to the pages we read.