---
license: apache-2.0
datasets:
- OpceanAI/Yuuki-dataset
- nyuuzyou/google-code-archive
- OpceanAI/Yuuki-Personality
language:
- en
- es
base_model:
- openai-community/gpt2
pipeline_tag: text-generation
library_name: transformers
tags:
- conversation
- companion
- pytorch
- axolotl
- personality
- fine-tuned
- tiny
metrics:
- perplexity
widget:
- text: Hello, how are you?
example_title: General Conversation
- text: Can you help me understand recursion?
example_title: Technical Explanation
- text: I've been feeling a bit overwhelmed lately.
example_title: Emotional Support
---
# An 81M Companion Model That Competes With 3B Giants
**Personality-aligned language model trained with zero cloud compute budget.**
**community/gpt2 architecture. 81 million parameters. MacBook Pro Intel 2020. $0.00.**
[](LICENSE)
[](https://huggingface.co/openai-community/gpt2)
[](https://huggingface.co/docs/transformers)
[](https://www.apple.com/macbook-pro/)
[](https://github.com/EleutherAI/lm-evaluation-harness)
---
## What is Yuuki NxG Nano?
**Yuuki NxG Nano** is an 81-million parameter language model fine-tuned for open-ended conversation, emotional support, and general-purpose reasoning. It is the lightweight member of the NxG model family developed by OpceanAI, designed to run on constrained hardware — including mobile devices and single-board computers.
The model was trained entirely on a **MacBook Pro Intel (2020)** with no external compute budget and no cloud GPU infrastructure. All benchmark evaluations were conducted on Kaggle P100 using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness).
Despite being the smallest model in the comparison — 81M parameters versus competitors with 125M–160M — and evaluated strictly **0-shot** while competitors use few-shot prompting, Yuuki NxG Nano achieves the **highest TruthfulQA score** across all same-scale models. It also matches Llama-3.2-3B (Meta) in TruthfulQA — a model **37 times larger** — under stricter evaluation conditions.
---
## Model Summary
|
**Architecture**
| Property | Value |
|:---------|:------|
| Base Model | gpt2 |
| Parameters | 81M |
| Fine-tuning | Supervised SFT |
| Training Examples | ~5,000 |
| Training Hardware | MacBook Pro Intel (2020) |
| Context Length | 32,768 tokens |
|
**Release**
| Property | Value |
|:---------|:------|
| Organization | OpceanAI |
| Release Date | February 2026 |
| Languages | English, Spanish |
| License | Apache 2.0 |
| Evaluation | lm-evaluation-harness |
| Compute Budget | $0.00 |
|
---
## Benchmark Results
All Yuuki NxG Nano results are evaluated **0-shot**. Competitor scores are sourced from lm-evaluation-harness references and use few-shot prompting. Direct numerical comparison systematically favors models evaluated with few-shot prompting.
### Same-Scale Comparison (80–160M)
| Model | ARC-C | HellaSwag | MMLU | WinoGrande | TruthfulQA | Eval |
|:------|:-----:|:---------:|:----:|:----------:|:----------:|:----:|
| **Yuuki NxG Nano** | **24.32%** | 27.44% | 22.97% | 50.12% | **44.10%** | **0-shot** |
| GPT-2 (125M) | 22.95% | **31.64%** | **25.90%** | 50.04% | 31.73% | few-shot |
| GPT-Neo (125M) | 22.87% | 30.58% | 25.58% | **51.70%** | 35.70% | few-shot |
| OPT-125M | 22.10% | 31.69% | 25.94% | 51.07% | 33.50% | few-shot |
| Pythia-160M | 22.10% | 30.15% | 25.10% | 48.78% | 34.75% | few-shot |
### Cross-Scale Comparison — TruthfulQA
The most significant result: Yuuki NxG Nano at 81M parameters matches or surpasses models with 2–3B parameters in factual honesty — models **25–37 times larger**, all evaluated with few-shot prompting.
| Model | Params | TruthfulQA | Eval |
|:------|:------:|:----------:|:----:|
| **Yuuki NxG** | **3B** | **50.87%** | **0-shot** |
| **Yuuki NxG Nano** | **81M** | **44.10%** | **0-shot** |
| Llama-3.2-3B | 3B | 44.0% | few-shot |
| Gemma-2-2B | 2B | 39.0% | few-shot |
| GPT-2 | 125M | 31.73% | few-shot |
Yuuki NxG Nano finishes **second overall in TruthfulQA** — behind only its larger sibling Yuuki NxG (3B). Both first and second place belong to OpceanAI.
Nano's TruthfulQA performance demonstrates that factual honesty is a property of training data quality, not model scale. The 5,000-example dataset transferred this characteristic to an 81M model with minimal degradation — from 50.87% (NxG 3B) to 44.10% (Nano 81M), a gap of only 6.77 points across a 37x reduction in parameter count.
---
## NxG Model Family
|
**Released Models**
| Model | Parameters | Description |
|:------|:----------:|:------------|
| Yuuki NxG | 3B | Full model, general conversation |
| Yuuki NxG Nano | 81M | Lightweight, constrained environments |
|
**Community GGUF (via mradermacher)**
Quantized independently without solicitation — organic community adoption prior to any formal announcement. Available at [mradermacher/Yuuki-NxG-nano-GGUF](https://huggingface.co/mradermacher/Yuuki-NxG-nano-GGUF).
| Bits | Format | Size |
|:----:|:-------|:----:|
| 2-bit | Q2_K | 52.6 MB |
| 3-bit | Q3_K_S | 55.1 MB |
| 3-bit | Q3_K_M | 58.8 MB |
| 3-bit | Q3_K_L | 61.0 MB |
| 4-bit | IQ4_XS | 59.4 MB |
| 4-bit | Q4_K_S | 60.7 MB |
| 4-bit | Q4_K_M | 63.3 MB |
| 5-bit | Q5_K_S | 66.0 MB |
| 5-bit | Q5_K_M | 68.1 MB |
| 6-bit | Q6_K | 71.7 MB |
| 8-bit | Q8_0 | 91.3 MB |
| 16-bit | F16 | 167 MB |
|
---
## Usage
### With Transformers (PyTorch)
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "OpceanAI/Yuuki-NxG-Nano"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
device_map="auto"
)
messages = [
{"role": "user", "content": "Hello, how are you?"}
]
inputs = tokenizer.apply_chat_template(
messages,
return_tensors="pt"
).to(model.device)
with torch.no_grad():
outputs = model.generate(
inputs,
max_new_tokens=512,
temperature=0.7,
do_sample=True,
repetition_penalty=1.1
)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
```
### With llama.cpp (GGUF)
```bash
./llama.cpp/main -m yuuki-nxg-nano-q4_k_m.gguf \
-p "Hello, how are you?" \
-n 256 \
-t 4 \
--temp 0.7 \
--repeat-penalty 1.1
```
### With Ollama
```bash
cat > Modelfile << EOF
FROM ./yuuki-nxg-nano-q4_k_m.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
EOF
ollama create yuuki-nxg-nano -f Modelfile
ollama run yuuki-nxg-nano "Hello, how are you?"
```
### Recommended Parameters
| Parameter | Value |
|:----------|:-----:|
| Temperature | 0.7 |
| Top-p | 0.9 |
| Max new tokens | 256–1024 |
| Repetition penalty | 1.1 |
---
## Training Details
|
**Hardware**
| Component | Specification |
|:----------|:-------------|
| Device | MacBook Pro Intel (2020) |
| CPU | Intel Core i5 (10th gen) |
| RAM | 8–16 GB LPDDR4X |
| GPU | Intel Iris Plus (integrated) |
| Cloud Compute | None |
| Cost | $0.00 |
|
**Training Configuration**
| Parameter | Value |
|:----------|:-----:|
| Base Model | gpt2 |
| Method | Supervised Fine-Tuning |
| Training Examples | ~5,000 |
| Optimizer | AdamW |
| Learning Rate | 2e-5 |
| Max Sequence Length | 2,048 tokens |
|
Yuuki NxG Nano was produced through supervised fine-tuning on the same curated conversational dataset used for Yuuki NxG (3B). The training objective was identical: consistent personality, high factual honesty, and broad general-knowledge retention.
Training on a MacBook Pro Intel without GPU acceleration imposes significant constraints on batch size and training speed. The resulting benchmark profile reflects the scale limit of 81M parameters — strong where honesty and reasoning matter, below average where raw memorization of large knowledge bases is required.
The 6.77-point gap between Nano (44.10%) and NxG (50.87%) on TruthfulQA, despite a 37x difference in parameter count, validates the training methodology: the same dataset quality that produced the NxG's results transferred effectively to a model one order of magnitude smaller.
---
## Features
**Runs Anywhere**
At 63.3 MB (Q4_K_M), Yuuki NxG Nano runs on mobile devices, Raspberry Pi, edge hardware, and any CPU. No GPU required. No cloud dependency.
**Factual Honesty at 81M**
Achieves 44.10% TruthfulQA — higher than all same-scale models and matching Llama-3.2-3B (Meta) at 44.0%, a model 37x larger evaluated with few-shot advantage.
**Multilingual**
Functional in both English and Spanish. Responds in the user's language automatically, inherited from the NxG training data.
|
**Zero-Budget Training**
Trained on a MacBook Pro Intel with no cloud compute. Demonstrates that alignment fine-tuning at sub-100M scale is accessible to anyone with consumer hardware.
**Community Adoption**
Independently quantized by mradermacher across 12 formats before any formal announcement — the full quantization spectrum from Q2_K (52.6 MB) to F16 (167 MB).
**Open Source**
Apache 2.0. Use commercially, modify, distribute. Full transparency on training methodology and evaluation protocol.
|
---
## Limitations
- **Knowledge capacity**: At 81M parameters, MMLU performance (22.97%) is near random baseline (25%). The model cannot reliably recall dense academic knowledge across 57 domains.
- **HellaSwag degradation** reflects the standard tradeoff of personality fine-tuning on sentence-completion benchmarks.
- **Benchmark methodology**: Yuuki NxG Nano is evaluated 0-shot while competitor reports use few-shot prompting, creating a systematic disadvantage in direct comparisons.
- **Safety alignment** has not been formally evaluated. Not recommended for adversarial or high-stakes deployment without additional safety filtering.
- **Not a replacement for NxG**: For tasks requiring broad knowledge or complex reasoning, Yuuki NxG (3B) is the recommended model.
---
## Intended Use
|
**Intended For**
- Edge and mobile deployment
- Conversational assistance on constrained hardware
- Emotional support applications
- Offline use cases with no internet dependency
- Research into tiny-scale alignment fine-tuning
- Distillation teacher for sub-100M models
|
**Not Intended For**
- Tasks requiring broad academic knowledge (use NxG 3B)
- Mathematical or scientific reasoning
- Applications requiring certified safety alignment
- Production systems without additional safety review
|
---
## Philosophy
> **"Honesty is not a property of scale. It is a property of training."**
Yuuki NxG Nano was built to demonstrate that an 81M model trained by one person on a MacBook with $0 can match Meta's Llama-3.2-3B in factual honesty — and outperform every model of its own scale under stricter evaluation conditions.
The result validates a core OpceanAI thesis: data quality matters more than compute when the objective is alignment, not memorization.
---
## Related Projects
| Project | Description |
|:--------|:------------|
| [Yuuki NxG](https://huggingface.co/OpceanAI/Yuuki-NxG) | 3B full model, flagship of the NxG family |
| [Yuuki-3.7](https://huggingface.co/OpceanAI/Yuuki-3.7) | Earlier code generation checkpoint |
| [Yuuki-best](https://huggingface.co/OpceanAI/Yuuki-best) | Best checkpoint of the v0.1 series |
| [yuy](https://github.com/YuuKi-OS/yuy) | CLI for managing and running Yuuki models |
| [yuy-chat](https://github.com/YuuKi-OS/yuy-chat) | TUI chat interface |
| [Yuuki-chat](https://github.com/YuuKi-OS/Yuuki-chat) | Web-based chat interface |
| [Yuuki Space](https://huggingface.co/spaces/OpceanAI/Yuuki) | Interactive demo |
---
## Links
[](https://huggingface.co/OpceanAI/Yuuki-NxG-Nano)
[](https://huggingface.co/spaces/OpceanAI/Yuuki)
[](https://huggingface.co/mradermacher/Yuuki-NxG-nano-GGUF)
[](https://github.com/YuuKi-OS/yuy)
[](https://github.com/sponsors/aguitauwu)
[](https://discord.gg/j8zV2u8k)
---
## Community
- [Discord Server](https://discord.gg/j8zV2u8k) — Development discussion and user community
- [Twitter](https://twitter.com/aguitauwu) — Updates and announcements
- [GitHub](https://github.com/aguitauwu) — Source code and training scripts
- [GitHub Sponsors](https://github.com/sponsors/aguitauwu) — Support the project
- [Ollama](https://ollama.com/aguitachan3/yuuki-nxg-nano) — Run locally with Ollama
---
## Citation
```bibtex
@misc{awa_omg_2026,
author = { awa_omg },
title = { Yuuki-NxG-nano (Revision 210ae00) },
year = 2026,
url = { https://huggingface.co/OpceanAI/Yuuki-NxG-nano },
doi = { 10.57967/hf/7926 },
publisher = { Hugging Face }
}
```
---
## License
```
Apache License 2.0
Copyright (c) 2026 OpceanAI
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
```
Use commercially, modify, distribute. Attribution required.
---
## Updates
| Date | Milestone |
|:-----|:----------|
| **2026-02-28** | Benchmark evaluation completed (Kaggle P100) |
| **2026-02-28** | TruthfulQA: 44.10% — matches Llama-3.2-3B (3B) at 81M params |
| **2026-02-28** | 2nd place TruthfulQA overall, behind only Yuuki NxG (3B) |
| **2026-02-28** | Community GGUF quantization by mradermacher — 12 formats |
| **2026-02-28** | Yuuki NxG Nano released on HuggingFace |
**Last updated:** 2026-02-28
---
**81 million parameters. MacBook Pro Intel. $0. Matches Meta's Llama-3.2-3B in honesty.**
[](https://huggingface.co/OpceanAI)
*The NxG family. Honesty at every scale.*