NeuralAI · An independent AI lab · West Memphis, Arkansas
Trained from scratch. Every single weight owned.
NEA1 is a 100M-parameter language model designed, tokenized, and trained entirely in-house — no borrowed weights, no API wrappers, no borrowed identity. It chats live below, holds a conversation, and calls tools mid-answer.
Live model
Talk to NEA1 — right here
This is the real model on NeuralAI's own inference stack, streaming token by token. Ask it a question, follow up on its answer, or give it a page to fetch. First reply can take a few seconds while the backend wakes.
Ask me anything — then follow up. I remember the conversation, switch topics with you, and I'll fetch a page or tell you plainly when a tool isn't available.
The model
NEA1, on the record
A 12-layer Llama-style transformer with a tokenizer we built ourselves — released with its model card and evaluation results.
Built, not assembled. NEA1's architecture, tokenizer, and training pipeline were written from first principles. The base model pretrained on 768M FineWeb tokens; the SFT line then taught it conversation, follow-ups, recall, and tool use — one version at a time, each gated on published evals.
Small on purpose. At 100M parameters NEA1 runs on commodity CPUs with room to spare. It is a research instrument and a proof of sovereignty: a lab this size can own its whole stack.
Open. Weights, tokenizer, and eval reports are public under Apache 2.0 on Hugging Face and GitHub.
Benchmarks
Measured, not marketed
Every SFT version ships only after clearing the same gate. These are the serving checkpoint's numbers, exactly as evaluated.
Roadmap
Where NEA1 is
SFT v13.2
The date line, closed: full month-boundary tomorrow rolls and honest no-date declines — behind the same release gate. Serving above and on the full chat.
SFT v14
The Awareness update: source awareness, calibrated uncertainty, and live data streams — behind the same release gate.
SFT v12.2
The previous production checkpoint. Superseded by v13.2; kept public as the prior release.
Company
An independent lab, on purpose
NeuralAI was founded by DeAndrew Preston Harris in West Memphis, Arkansas, with a conviction that has not moved: AI sovereignty starts with owning the fundamentals — the architecture, the tokenizer, the data, and the weights — rather than renting intelligence through someone else's API.
NEA1 is the lab's first model and its public proof of work: small, open, and honest about what it can and cannot do.
Contact
Work with the lab
Research partnerships, model access, and general inquiries.