Small models are easier to understand.
We can read the code, follow the memory, and see where a small model fails. We make a model larger only after a smaller test shows that the idea works.
SMALL MODELS · CLEAR TESTS · LOCAL AI
Build useful AI
without hiding how it works.
We build small AI models that we can inspect, test, and run on a personal computer. ZERO.4 is our main public model. We also study language, memory, checked tools, integer models, and teams of small expert models.
WHAT THE LAB IS
A model is only one part of the system.
A model can learn language. Memory can keep useful parts of a conversation. Exact programs can check answers. When the system does not know, it should say so.
Your words can stay on your computer.
ZERO.4 runs inside the browser. It does not send your prompt to a server. There is no hidden model or outside service fixing its answers.
Good writing is not proof.
The language model writes. Memory recalls. A router chooses a tool. Exact code checks facts it can prove. We keep these jobs separate and test each one.
Failed tests still teach us.
We keep failed experiments. They show which path did not work and help us ask a better question. ilXyr keeps the plan, result, model file, and decision for each test.
ONE PROGRAM, MANY BRANCHES
The projects are parts of one lab.
Each project has one job. Good results feed the next ZERO model. Failed results stay in the record.
ZERO
Our public model family. It includes the small language model, browser app, conversation format, and model history.
Sero
Our language training branch. It tests training data, token formats, model size, and new ways to group bytes into tokens.
Holo
Our memory branch. It tests how a small model can keep useful old moments without saving the full conversation.
Solomon / NSRL
Our integer-model branch. It trains models that can replay the same way, run well on CPUs, and work with text and images.
CRLP
Our checking branch. It builds small task models and exact tests that decide whether a research claim is ready to publish.
ilXyr
Our research record. It freezes the test plan, limits the budget, tracks model files, records repeats, and stores the final decision.
THE NEXT MODEL FAMILIES
Bring the work together before making it larger.
The family numbers mark research steps. They do not mean a fixed model size. A family becomes public only after it passes a test we wrote before training.
BETTER LOCAL LANGUAGE
A stronger model that still runs locally.
Can a small local model write and talk much better while staying easy to inspect?
Z5 brings together ZERO.4, Sero's approved training data, its current token format, larger language training, conversation memory, and Holo recall. The public result must still run on a personal computer.
MUST PASS
- Beat ZERO.4 on fixed language and conversation tests.
- Show that limited memory helps.
- Pass the same planned test with several random seeds.
- Publish the chosen model file and its limits.
LANGUAGE WITH CHECKED TOOLS
A model that knows when to use a tool.
Can the language model choose tools and memories while exact code still checks the final result?
Z6 brings together Z5, ZERO's tool-routing work, CRLP's checks, ilXyr's test plans, and a matching Solomon integer model. The model may choose an operation. Exact code performs and checks it.
MUST PASS
- Lock the shared tasks before training.
- Count wrong tool choices and safe refusals.
- Compare every test case, not only the average.
- Repeat the result with the integer model.
A TEAM OF SMALL EXPERTS
Many experts in one clear system.
Can small expert models, memory, text, and images work together without hiding which part did the work?
Z8 brings together Z6, Holo memory, Solomon text-and-image models, groups of routed experts, and small task scorers. We add one expert at a time. Each expert must prove that it helps.
MUST PASS
- Keep the Z5 language and Z6 safety results.
- Remove each expert once to prove that it helps.
- Pass new text, image, and task tests.
- Publish memory use, speed, and energy use.
RESEARCH CONSOLIDATION
One model history. Clear owners. Fewer copies.
Each project can keep its own code. One catalog should show every model, who owns it, where its file is, whether it passed, and which test supports the result.
- Keep this repository as the main public ZERO model and browser app.
- Give every model an ID, a parent model, a file hash, a status, and a test record.
- Store the official result in ilXyr. Keep training code in the project that owns it.
- Mark old work as a Z5 input, Z6 input, Z8 input, product test, or archive.
- Freeze copied repositories after we save any unique results they contain.
- Every new experiment must say which Z5, Z6, or Z8 decision it could change.