# nagisanzenin: nagi

> Nagi, an independent developer's decision family: LoRA adapters (and a small encoder) that score a closed set of 2 to 26 supplied options in one forward pass and return a choice with a probability per option, with no generated reasoning. This page carries the 27B adapter.

- Page: https://systemonemodels.ai/nagisanzenin/nagi
- API: https://api.systemonemodels.ai/v1/models/nagisanzenin/nagi
- Download: `pip install systemonemodels && systemone pull nagisanzenin/nagi`

## Facts

| | |
|---|---|
| Maker | nagisanzenin (https://systemonemodels.ai/nagisanzenin) |
| Decides | choice, score, noul |
| Architecture | nagi |
| Base model | qwen/qwen3.8-27b |
| Parameters | 27B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-25 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: Nagi public four-system decision suite (8 public sources adapted to typed choice, 4,518 rows, common evidence), the maker's own run. Numbers are the publisher's own.

- Decision accuracy: 82.8%

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# Nagi-ENORMOUS (27B)

Nagi, an independent developer's decision family: LoRA adapters (and a small encoder) that score a closed set of 2 to 26 supplied options in one forward pass and return a choice with a probability per option, with no generated reasoning. This page carries the 27B adapter.

A personal project. The SDK's system_one(state, questions) takes choice, score and noul questions. ENORMOUS is a LoRA adapter on Qwen3.8-27B, so the base is needed. Siblings: Nagi-HUGE (Gemma 4 12B LoRA), nagi-big-v3 (Qwen3.5-4B; its card declares no licence), nagi-smol-v0 (ModernBERT-large, a pickle checkpoint) and ENORMOUS-Burst. On the maker's own suite adapting eight public sources to typed choice (ANLI, XNLI, BoolQ, CB, COPA, RTE, WiC, MMLU-Pro) ENORMOUS scores 82.76%; the maker notes exposure to these sources during pretraining is unknown and that none are official leaderboard scores. Option order changes some answers (9.7 to 13.1% flips for the smaller sizes). The maker's own game exhibitions against Jev are reported on the card.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 27B |
| Base model | `qwen/qwen3.8-27b` |
| Maker | nagisanzenin |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 82.8% |
| Reported latency | 122 ms at 1,024 tokens with the adapter merged and fused kernels |

## Get the weights

```bash
pip install systemonemodels
systemone pull nagisanzenin/nagi
```

The files are served from the maker's Hugging Face repository, [`nagisanzeninz/Nagi-ENORMOUS`](https://huggingface.co/nagisanzeninz/Nagi-ENORMOUS), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/nagisanzeninz/Nagi-ENORMOUS)
- [SDK and docs](https://github.com/nagisanzenin/nagi)
- [Nagi-HUGE](https://huggingface.co/nagisanzeninz/Nagi-HUGE)

---

*This page was opened by System One for nagisanzenin, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models