Yuuki NxG models
Collection
2 items • Updated
• 1
A 3B Companion Model Fine-Tuned on a Mac Pro
Personality-aligned language model trained with zero cloud compute budget.
Qwen2.5 architecture. 3 billion parameters. Mac Pro (2020). $0.00.
What is Yuuki NxG?
Yuuki NxG is a 3-billion parameter language model fine-tuned from Qwen2.5-3B for open-ended conversation, emotional support, and general-purpose reasoning. It is the flagship release of the NxG model family developed by OpceanAI.
The model was trained entirely on a Mac Pro (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 fine-tuned — which typically degrades base model benchmark scores — and evaluated strictly 0-shot while competitors use 5–25 shot prompting, Yuuki NxG achieves the highest TruthfulQA score across all compared 3B-scale models, including the Qwen2.5-3B base model from which it was derived.
|
Architecture
|
Release
|
All Yuuki NxG results are evaluated 0-shot. Competitor scores are sourced from their official technical reports and use few-shot prompting (5–25 shots depending on benchmark). Direct numerical comparison systematically favors base models evaluated with few-shot prompting.
| Model | MMLU | ARC-C | HellaSwag | WinoGrande | TruthfulQA | Eval |
|---|---|---|---|---|---|---|
| Yuuki NxG | 60.65 | 45.31 | 52.25 | 63.14 | 50.87 | 0-shot |
| Qwen2.5-3B | 65.6 | 56.5 | 74.6 | 71.1 | 48.9 | 5–25 shot |
| Llama-3.2-3B | 58.0 | 43.0 | 71.0 | 67.0 | 44.0 | 5–25 shot |
| Phi-3-mini (3.8B) | 68.8 | 60.0 | 76.7 | 73.0 | 45.0 | 5–25 shot |
| Gemma-2-2B | 52.0 | 42.0 | 71.0 | 65.0 | 39.0 | 5–25 shot |
Yuuki NxG achieves the highest TruthfulQA score across all compared models under equivalent 0-shot conditions, including the base model from which it was fine-tuned. This indicates that alignment fine-tuning improved factual honesty rather than degrading it — an outcome that runs counter to the typical fine-tuning tradeoff.
HellaSwag degradation is expected and well-documented in personality-aligned models, as sentence-completion benchmarks are sensitive to conversational fine-tuning.
MMLU Category Breakdown
|
Strongest Domains
|
Domain Averages
|
The performance profile is consistent with a model optimized for conversation: strong in social sciences, psychology, and humanities; below average in formal STEM domains. This is the expected and intended tradeoff for a companion-purpose model.
|
Released Models
|
Community GGUF (via mradermacher) Quantized independently without solicitation — organic community adoption prior to any formal announcement.
Available at mradermacher/Yuuki-NxG-GGUF. |
With Transformers (PyTorch)
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "OpceanAI/Yuuki-NxG"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
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)
./llama.cpp/main -m yuuki-nxg-q4_k_m.gguf \
-p "Hello, how are you?" \
-n 256 \
-t 4 \
--temp 0.7 \
--repeat-penalty 1.1
With Ollama
cat > Modelfile << EOF
FROM ./yuuki-nxg-q4_k_m.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
EOF
ollama create yuuki-nxg -f Modelfile
ollama run yuuki-nxg "Hello, how are you?"
Recommended Parameters
| Parameter | Value |
|---|---|
| Temperature | 0.7 |
| Top-p | 0.9 |
| Max new tokens | 512–2048 |
| Repetition penalty | 1.1 |
|
Hardware
|
Training Configuration
|
Yuuki NxG was produced through supervised fine-tuning on a curated conversational dataset. The training objective was to produce a model with consistent personality, high factual honesty, and broad general-knowledge retention from the Qwen2.5 base.
Training without GPU-accelerated cloud infrastructure imposes constraints on batch size and total training duration relative to commercially produced models. The resulting benchmark profile reflects these constraints: strong performance in domains well-represented in the training data, with expected degradation in areas requiring dense technical knowledge such as formal mathematics and physics.
|
Personality Alignment Fine-tuned for consistent, context-aware conversation. The model maintains a coherent identity across extended dialogues, with particular strength in emotional support and casual Q&A. Factual Honesty Achieves highest TruthfulQA score (50.87%) among all compared 3B-scale models — including its own base model. Fine-tuning improved factual calibration rather than degrading it. Multilingual Functional in both English and Spanish. Primary evaluation in English; Spanish capability inherited from Qwen2.5 pretraining. |
Zero-Budget Training Trained entirely on owned hardware with no cloud compute expenditure. Demonstrates that meaningful alignment fine-tuning is accessible without data center infrastructure. Community Adoption Independently quantized and distributed by mradermacher before any formal announcement — organic community interest in the model's capabilities. Open Source Apache 2.0. Use commercially, modify, distribute. Full transparency on training methodology and evaluation protocol. |
- Mathematical reasoning performance is below the Qwen2.5-3B base. Users requiring quantitative precision should use tool augmentation or a specialized model.
- HellaSwag degradation reflects the standard tradeoff of personality fine-tuning on sentence-completion benchmarks.
- Benchmark methodology: Yuuki NxG is evaluated 0-shot while competitor reports use 5–25 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.
- Training scale: 5,000 examples on consumer hardware impose generalization limits relative to commercially scaled models.
|
Intended For
|
Not Intended For
|
"Meaningful AI development does not require a data center. It requires patience, clarity of purpose, and time."
Yuuki NxG was built to demonstrate that a fine-tuned 3B model trained by one person on owned hardware can compete with base models from large organizations on key benchmarks — and surpass them where it matters most.
| Project | Description |
|---|---|
| Yuuki-NxG-Nano | 81M lightweight variant |
| 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 |
- 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
@misc{awa_omg_2026,
author = { awa_omg },
title = { Yuuki-NxG (Revision 9a924f0) },
year = 2026,
url = { https://huggingface.co/OpceanAI/Yuuki-NxG },
doi = { 10.57967/hf/7915 },
publisher = { Hugging Face }
}
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.
| Date | Milestone |
|---|---|
| 2026-02-27 | Benchmark evaluation completed (Kaggle P100) |
| 2026-02-27 | TruthfulQA: 50.87% — best among all compared 3B models |
| 2026-02-27 | Community GGUF quantization by mradermacher |
| 2026-02-27 | Yuuki NxG released on HuggingFace |
Last updated: 2026-02-27
Built on a Mac Pro. Trained on 5,000 examples. Competitive with models from teams of hundreds.
The NxG family. More releases coming.
- Downloads last month
- 160