The models behind nenspace
nenspace trains its own models. This page is the record: what trained, on what data, filtered how, measured against what. Nothing here is licensed marketing copy; it is the logbook of the training runs.
#1 · nen v2.0 · the default
The company thesis in weights: as useful as the frontier assistants, none of the excess language. Answers the question asked, commits, says what is not known, revises on new information.
trained · 2026-08-23 · 36 minutes · about $4.77
data · 510 rows · each read by a person before it trained · zero user conversations, by construction
filtered · 878 candidates judged down to 510 · register floors enforced in code: no em dashes, no first person, no padding
measured · internal bench, n=95, same judge for every model
Answers in full: 100% (the same base under a prompt alone: 83.2%). Commits to a position: 81.8% (about 44%). Says what is not known: 100%. First person leaks in 2.1% of replies.
disclosed regression · noticing what a question needs: 55.6% vs 73.5% prompted · the next corpus's first job
Text dialogue only, on purpose. Web, tools and image input stay off until they are trained in: nothing is switched on at inference that the voice was not trained on.
The base model stays unnamed. The work is the register, and the register lives in the weights, not in a prompt.
#2 · nen-1 · the dialogue option
The register's first life. Question-led, median twenty words, trained on 395 voice rows. Where v2.0 answers, nen-1 listens and hands the question back. Kept, by choice, for conversations that want a mirror held at arm's length rather than an answer.
trained · 2026 · 395 rows · the same register floors
serves · the dialogue option in the model picker
Every training row is read by a person before it trains. No data is sold. nenspace is entirely self-funded.
nenspace · 2026-08