You can deploy this model and start earning money today!

Views

No views yet

An 81M Companion Model That Competes With 3B Giants

Name: Yuuki-NxG-nano
Availability: InStock

Personality-aligned language model trained with zero cloud compute budget.
community/gpt2 architecture. 81 million parameters. MacBook Pro Intel 2020. $0.00.

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.

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.

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	3B full model, flagship of the NxG family
Yuuki-3.7	Earlier code generation checkpoint
Yuuki-best	Best checkpoint of the v0.1 series
yuy	CLI for managing and running Yuuki models
yuy-chat	TUI chat interface
Yuuki-chat	Web-based chat interface
Yuuki Space	Interactive demo

Links

Community

Discord Server — Development discussion and user community
Twitter — Updates and announcements
GitHub — Source code and training scripts
GitHub Sponsors — Support the project
Ollama — 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.

The NxG family. Honesty at every scale.