A model for everyone, built from scratch.
Creating AI for everyone, from scratch. Centralia is an independent lab training its own language model, MDean, for people, not for big labs or experts only. Step by step, in public, with the numbers on the page. Join the waitlist, follow the run, or open the investor and grant briefing.
Be first to try MDean and follow each training step as it lands.
Honest eval gates at each step: a larger model is only trained when the smaller one finishes and its evaluations say the system is ready. Results are reported as measured, including where they fall short.
Safety guidance is part of the training material itself, so expected behavior is learned during training, not patched on afterward.
The model is built around eight principles covering truth, harm, fairness, and accountability, the standards its answers are measured against.
MDean is being taught to work with documents: read a file, summarize it, and work on its contents, with the same safety principles applied to what it reads and writes.
The build has been shared openly all along, as it happened. Follow the trail, or support the next step:
- About five months agoThe idea and the early work begin, months before any repo exists.
- Lab foundedCentralia Cognition Labs founded and the repo created.
- Spectrum Cycle chosenThe nine state ring becomes the lab identity.
- Building in public beginsProgress posted openly as the work happens.
- MDean training launchesThe 1B run starts.
- 1 billion tokens crossedFirst billion tokens of the run complete.
- 50 percent, then past 68Halfway November 5, 60.09 percent November 6, 68.70 percent November 7.
- MDean finishesExpected completion, Wednesday, November 11.
Every part of this was built by one solo founder: the tokenizer, the data, the training system, the safety work, and the public updates. That is why five months is the honest timeline, and it is exactly what funding changes. Funding buys H100 compute and help, so the 3B step and everything after it moves at the speed of the work, not the speed of one person and one machine.
I am Michael Smith. I taught myself this work, and I have ten years of building behind me. I did not wait for a lab to hire me or a team to hand me the hard parts. I built the lab.
Nobody handed me an education in this either. I learned it myself over ten years, through books, Instagram, Twitter, and YouTube, lesson by lesson, usually after work. School came later: I am a working adult now in my third year of a BS in Business Information Systems, with a career across several levels of technology behind me. I have also done contract work evaluating and training frontier AI models, the kind of work where you learn exactly how the best models are judged, and where they fall short.
Centralia is mine, built solo, from scratch: the custom tokenizer, the 19.5 billion token data diet, the training harness I wrote by hand, and MDean, the 1 billion parameter model that is 68.7 percent trained right now. Nobody handed me a stack. I am the stack.
I measure everything and I post the numbers as they are, good or bad. That is the whole pitch: I ship, I measure honestly, and I keep going. What comes next is simple. Finish MDean. Run the eval gates and publish what they say. Add the safety work and file handling. Then scale to the 3B.
— MICHAEL SMITH, FOUNDER, CENTRALIA COGNITION LABS
Tokenizer · 19.5B token data diet · hand built training harness · MDean, 1B, training now
Self taught over ten years through books, Instagram, Twitter, and YouTube · a working adult with a career across several levels of technology
BS in Business Information Systems, Grand Canyon University, third year, started September 2024 · Google Data Analytics certification, completed October 2025
Evaluating and training frontier AI models
Syndero and Code Sonar
Finish MDean · eval gates · safety and file handling · then the 3B
The cycle ends where it warms: amber, the waitlist. Early access, investor briefings, grant and funding conversations, and research partnerships all start here.
CENTRALIACOGNITIONLABS.COM · INVESTOR AND GRANT CONTACT: MICHAEL@CENTRALIACOGNITIONLABS.COM
All MDean figures are from the November 7, 2026 run record: step 13,112 of 19,074, 68.7 percent, 6.875B of 10B tokens, loss 2.9392, throughput 9,501 tok/s, expected finish Wednesday, November 11, 2026. The progress graphic shown on this page, in Figure 1 and again as the latest entry in the trail below, is the November 7, 2026 graphic, regenerated from that same record; the earlier trail entries are the November 2 monitor and the November 5 halfway graphic. Timeline dates are the verified milestones listed above; the conception point is deliberately undated, about five months before November 2026.