README.md

---
license: other
language:
  - en
tags:
  - legalbench
  - legal
  - evaluation
  - closed-weight
  - api
inference: false
---

# nonnon-small

NONNON-small is a closed-weight deep reasoning model served through an OpenAI-compatible API. It is designed for LegalBench-style tasks and is available for evaluation through `https://small.nonnon.ai/v1/chat/completions`.

## Access

Email labs@nonnon.ai to request an evaluation token. Tokens are free for legitimate evaluation purposes and rate-limited for fair use.

## How To Use

```python
from openai import OpenAI

client = OpenAI(
    api_key="YOUR_NONNON_TOKEN",
    base_url="https://small.nonnon.ai/v1",
)

response = client.chat.completions.create(
    model="nonnon1-small-router",
    messages=[{"role": "user", "content": "Your LegalBench prompt"}],
    temperature=0,
)

print(response.choices[0].message.content)
```

## Public Reproduction

GitHub reproduction recipe: https://github.com/NonnonLabs/nonnon-small-legalbench

Python SDK: https://github.com/NonnonLabs/nonnon-sdk-python

## Evaluation

Public validation runs on 2026-05-21:

| Run | Rows | Tasks | Macro | Micro |
|---|---:|---:|---:|---:|
| One row per task | 162 | 162 | 0.9862 | 0.9862 |
| Stratified 50 per task | 8,088 | 162 | 0.9506 | 0.9507 |

Independent evaluation is pending.

## Disclosure

This Hugging Face repository is a model card only. It intentionally contains no weights, no training data, no deployment scripts, and no private infrastructure details.

For questions, email labs@nonnon.ai.