Three Cell Node
Every module shares one internal shape: a Knowledge cell, a Thinking cell and a Reasoning cell.
Read about 3CNYaob / A model family by Fylume
Intelligence, closer to home.
A proposed model family by Fylume, built on research into the 3CN & MPN architecture.
Interactive · illustrative simulation
See the routing idea
Type a question. The diagram shows which domain modules a simple router would select, and which stay asleep.
Illustrative simulation. The interactive demonstration needs JavaScript. It shows which domain modules a simple router would select for a question — for example:
- “Why did Napoleon lose at Waterloo?” → history, politics
- “Solve x² − 5x + 6 = 0” → math
- “How do mRNA vaccines train the immune system?” → biology, medicine
- “Why is my Python loop so slow?” → software, cs
- “Size a solar system for a small cabin” → elec, energy, math
What is Yaob?
Yaob is Fylume’s proposed AI model family and the lab’s first flagship initiative. It is a research vehicle for one idea: a small shared cognitive core plus specialised modules, activated selectively, could deliver useful capability on ordinary hardware.
Yaob has not been released. There are no published benchmarks yet.
- Efficient inference
- Local execution
- A shared cognitive core
- Modular specialised knowledge
- Selective activation
- Reduced redundant computation
- Configurable resource budgets
- User control
- Flexible future model sizes
- Competitive capability at lower computational cost — as a research question
How it works
How the architecture works
- A request reaches the shared core.
- A router estimates which domain modules are relevant.
- Only those modules participate; the rest stay inactive.
- Each module is a Three Cell Node: knowledge, thinking, reasoning.
- The core integrates their contributions into one answer.
One shared cognitive core
Yaob’s intended architecture uses one shared cognitive core responsible for general language and reasoning. The exploratory size target is approximately 1–10 billion parameters, subject to research and validation.
Specialised functionality is reached through domain-relevant network modules. These are not separate full conversational models running inside every node.
Modular Parameter Network
Specialised knowledge lives in domain modules. The shared core reaches only for the modules a question needs.
Read about MPNCell-Node Parameters
Parameters are grouped by cell and node so modules can be added, replaced, disabled or removed.
Read about CNP- Knowledge Cell
- Domain-relevant knowledge representation.
- Thinking Cell
- Task interpretation, knowledge processing and intermediate computation.
- Reasoning Cell
- Reasoning-related operations; coordinating and evaluating task-relevant information.
3CN Three Cell Node
A proposed three-cell structure for organising the conceptual nodes of a modular architecture.
Read moreMPN Modular Parameter Network
A selectively accessible, domain-organised parameter network that a shared core can draw on when a task needs it.
Read moreCNP Cell-Node Parameters
An exploratory scheme for how parameters are grouped inside cells and nodes so components can be added, replaced or removed.
Read moreDomain selection
Only the knowledge you need, awake.
Try a prompt below or select a module to inspect it. This is an architecture visualiser, not a running model.
Domain modules: Mathematics, Statistics, Computer Science, History, Politics, Literature, Linguistics, Law, Music Theory, Architecture, Physics, Astronomy, Astrophysics, Chemistry, Biology, Medicine, Meteorology, Geology, Earth Science, Geography, Economics, Finance, Software Engineering, Machine Learning, Computer Networking, Computer Engineering, Computer Architecture, Mechanical Engineering, Electrical Engineering, Energy, Aerodynamics, Aerospace Engineering.
Why efficiency
Most questions need a little of a lot of knowledge, not all of it at once. If Yaob can avoid waking parameters a request does not need, it could run in less memory, on cheaper hardware, closer to the person using it.
Research statusExploratory research
Research status
Status: in research. The architecture is being designed and tested. No model weights are available, no benchmark results are published, and no hardware requirements have been verified.
Known limitations and risks
- Routing mistakes may reduce answer quality.
- Cross-domain questions are hard to split cleanly.
- Selective activation does not automatically reduce memory use.
- A small core may limit general reasoning ability.
- Every claim needs measurement on real hardware before it is made.
Model releases
Releases & downloads
No Yaob release is available yet. When one is, it will appear here with version, files, checksums, hardware notes and licence.
Yaob
Fylume’s proposed model family. In research — no weights have been released.
- Family
- Yaob
- Parameters
- Core target ≈1–10B (exploratory)
- Architecture
- 3CN & MPN (proposed)
Not released. No files are available to download.
Code & documentation
Research code and documentation will be published on GitHub as they become ready.
GitHub (opens in a new tab)Documentation will be published with the first release.
Questions
Frequently asked
Can I download Yaob today?
Not yet. Yaob is a proposed model family under active research. There is no public release, and the Models page will only list files once they genuinely exist, with checksums and licences.
Is the demo a real model?
No. Until a Yaob inference endpoint is connected, the demo is an illustrative simulation of the routing idea: a transparent keyword router picks domain modules and the page explains why. Sample answers are written by Fylume and labelled as such.
Does every node contain its own chatbot?
No. One shared cognitive core is responsible for general language and reasoning. Domain modules are specialised parameter groups the core can draw on — not separate full conversational models.
How big will Yaob be?
The exploratory target for the shared core is roughly 1–10 billion parameters. That range is a research target, not a commitment, and will change as experiments report back.
Will it run on my laptop?
Local execution on ordinary hardware is a central goal. We will not publish hardware requirements until we have measured them on real machines.
Will Yaob be open source?
Open research and code are encouraged wherever appropriate. Licence terms will be stated on each release; until then, no licence is implied.
Is Yaob better than GPT or Claude?
There is no evidence for that claim, and we will not make it. Frontier models from OpenAI, Anthropic and others are extraordinary systems. Our question is different: how much useful capability can fit in a much smaller, selectively activated budget? Benchmarks will be published with methodology when they exist.
