Chat with NEA1

NEA1 is NeuralAI's own language model — designed, tokenized, and trained from scratch in West Memphis, Arkansas. It holds a real conversation: ask a question, then follow up, switch topics, or let it call a tool mid-answer.

100,111,872

parameters

32,000

tokenizer vocab

2,048 tokens

context

SFT v12.2

serving checkpoint

Hey, I'm NeuralAI — how can I help you today?

Replies stream in token by token. First reply can take a few seconds while the backend wakes.

NEA1 is a research model and can make mistakes. When a tool is unavailable, it will tell you rather than invent an answer.

On the record

About this checkpoint

SFT v12.2 (step 12,000) is the current serving checkpoint, gated before release on the lab's published evals:

Golden-40 tool-use eval35.0 / 40

Multi-turn conversation suite11 / 12

Fluency guardrail PPL (50-doc FineWeb)48.1

Weights & full eval reportsHugging Face

Full conversation memory — follow-ups, topic switches, and recall across the whole chat.
Tool use mid-reply — fetch pages and read files; if a tool is unavailable it says so instead of making something up.
Three decoding modes — Focused, Balanced, and Creative, in the top bar.

The lab behind NEA1: neuralai.wasmer.app · GitHub