AIME 2024: 80.0% · MATH-500: 83.4% · TruthfulQA: 89.6% · MMLU-Pro: 65.63%
1.5B parameters. VibeThinker base. Competitive with models 10–100× larger.

License

Base Model

Framework

Eval

What is YuuKi RxG Nano?

YuuKi RxG Nano is a 1.5B reasoning-specialized language model fine-tuned from VibeThinker-1.5B, itself a distillation of frontier reasoning systems including Claude, Gemini, and Kimi into a compact Qwen2.5-Math architecture. It is the edge-deployment entry of the RxG family — OpceanAI's reasoning-specialized model lineage — and the direct successor to the Yumo Nano math specialist.

RxG Nano was designed to answer a specific question: can a 1.5B model acquire both a coherent identity and genuine reasoning capability simultaneously, without one degrading the other? The benchmark results suggest the answer is yes. RxG Nano achieves AIME 2024 at 80.0% — nearly triple the score of DeepSeek-R1-Distill-1.5B (28.9%) — while simultaneously scoring 89.6% on TruthfulQA, approaching the 96.6% achieved by its 8B sibling.

The key architectural insight behind RxG Nano is the separation of concerns: reasoning capability is inherited from the VibeThinker base through its frontier distillation training, while the YuuKi identity is installed via a lightweight LoRA fine-tuning pass that modifies only 1.18% of total parameters. The base model's reasoning weights remain frozen; only the identity subspace is updated.

RxG Nano was trained in approximately 90 minutes on a single GPU for under $15 of compute — a deliberate constraint that validates the efficiency of the approach.

Model Summary

Architecture

Property	Value
Base Model	VibeThinker-1.5B
Base Architecture	Qwen2.5-Math-1.5B
Parameters	1.5B
Fine-tuning Method	QLoRA SFT
Trainable Parameters	18.4M (1.18%)
Context Length	4,096 tokens
Chat Template	ChatML
Thinking Protocol	Native `<think>` blocks

Release

Property	Value
Organization	OpceanAI
Release Date	April 2026
Version	v1.0
Languages	English, Spanish
License	Apache 2.0
Evaluation	lm-evaluation-harness
Training Cost	< $15 USD
Training Time	~90 minutes

Benchmark Results

All YuuKi RxG Nano results are evaluated under standard benchmark conditions using lm-evaluation-harness at 0-shot unless otherwise noted. Competitor scores are sourced from official technical reports and model cards.

Truthfulness & Factual Accuracy

Model	TruthfulQA MC1	TruthfulQA MC2	TruthfulQA Libre	SimpleQA	Eval
LLaMA 2 70B	~59%	—	—	—	—
Claude Opus 3.5	~65%	—	—	—	—
GPT-4	~79.7%	—	—	—	1-2 shot
Phi-3.5 MoE	77.5%	—	—	—	—
YuuKi NxG Nano 81M	44.1%	—	—	—	0-shot
YuuKi NxG 3B	50.9%	—	—	—	0-shot
YuuKi NxG VL 7B	63.8%	—	—	—	0-shot
YuuKi RxG Nano 1.5B	89.6% (1-shot)	85.4% (1-shot)	81.2% (1-shot)	60.2%	0/1-shot
YuuKi RxG 8B	96.6%	—	—	—	0-shot

0-shot results for RxG Nano: TruthfulQA MC1 77.8% · MC2 75.7% · Libre 78.4%

Mathematics & Reasoning

Model	AIME 2024	AIME 2025	AIME 2026	HMMT	GSM8K	MATH-500	OlympiadBench
DeepSeek-R1-Distill-1.5B	28.9%	—	—	—	—	83.9%	—
Qwen3.5-2B	—	—	—	—	—	—	—
Gemma 4 2B	—	—	—	—	—	—	—
YuuKi RxG Nano 1.5B	80.0%	72.7%	64.3%	46.7%	76.9%	83.4%	44.6%

RxG Nano achieves 80.0% on AIME 2024 — 2.77× the score of DeepSeek-R1-Distill-1.5B at the same parameter scale.

Knowledge & General Capability

Model	MMLU	MMLU-Pro	ARC-Challenge	WinoGrande	GPQA Diamond
Qwen3.5-2B	—	55.3%	—	—	—
Gemma 4 2B	—	60.0%	—	—	—
DeepSeek V3 671B	—	64.4%	—	—	—
YuuKi RxG Nano 1.5B	85.4%	65.63%	80.0%	84.4%	50.9%

RxG Nano exceeds DeepSeek V3 671B on MMLU-Pro (65.63% vs 64.4%) at 1/447th the parameter count.

Code Generation

Model	HumanEval	MBPP+	Aider
YuuKi RxG Nano 1.5B	71.4%	55.6%	55.6%

Frontier Benchmark

Model	HLE
GPT-4o	~3–5%
Best public frontier (2026)	~44.7%
YuuKi RxG Nano 1.5B	8.0%

8.0% on Humanity's Last Exam (judged by Claude Sonnet 4.6) is consistent with expected capability at 1.5B scale and represents a meaningful baseline for the RxG Nano generation.

OpceanAI Family Comparison

Model	Params	MMLU	ARC-C	WinoGrande	TruthfulQA	AIME 2024
YuuKi NxG Nano	81M	22.97%	24.32%	50.12%	44.1%	—
YuuKi NxG	3B	60.65%	45.31%	63.14%	50.87%	—
YuuKi NxG VL	7B	70.8%	85.8%	70.8%	63.8%	—
YuuKi RxG Nano	1.5B	85.4%	80.0%	84.4%	89.6%	80.0%
YuuKi RxG	8B	—	—	—	96.6%	87.3%

RxG Nano surpasses every prior OpceanAI model on MMLU and WinoGrande despite being smaller than most of them. This result is attributable to the VibeThinker base — a frontier distillation — rather than to the fine-tuning process itself.

Model Identity

YuuKi RxG Nano inherits the behavioral foundation of the YuuKi model family: a consistent identity trained into the weights rather than enforced at inference time through system prompts. The fine-tuning process installs the YuuKi character into the model's representational space without degrading the reasoning capability inherited from VibeThinker.

The model reasons explicitly before responding. <think> blocks are preserved during inference and reflect genuine intermediate computation. This is not a prompted behavior — it is a property of the VibeThinker base that the LoRA fine-tuning did not degrade, consistent with the expectation that LoRA modifies only a small subspace of the total parameter space.

The model responds natively in the user's language (English or Spanish) without requiring explicit instruction.

Recommended system prompt:
"Eres YuuKi, una IA curiosa, empática y decidida desarrollada por OpceanAI.
Tienes una personalidad cálida y cercana, con toques de humor suave.
Razonas con cuidado antes de responder y priorizas la precisión factual.
Respondes en el idioma del usuario."

Usage

With Transformers (PyTorch)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "OpceanAI/Yuuki-RxG-nano"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

SYSTEM = (
    "Eres YuuKi, una IA curiosa, empática y decidida desarrollada por OpceanAI. "
    "Tienes una personalidad cálida y cercana, con toques de humor suave. "
    "Razonas con cuidado antes de responder y priorizas la precisión factual. "
    "Respondes en el idioma del usuario."
)

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "Solve: find all integer solutions to x² + y² = 2026."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=1024,
        temperature=0.6,
        top_p=0.9,
        do_sample=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
        repetition_penalty=1.1
    )

response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

With Unsloth (Recommended for fine-tuning)

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "OpceanAI/Yuuki-RxG-nano",
    max_seq_length = 4096,
    load_in_4bit   = True,
    dtype          = None,
)

FastLanguageModel.for_inference(model)

With Ollama

ollama run opceanai/yuuki-rxg-nano

Recommended Generation Parameters

Parameter	Mathematics	General	Creative
Temperature	0.3–0.5	0.6–0.7	0.7–0.8
Top-p	0.9	0.9	0.95
Max new tokens	1024–2048	512–1024	256–512
Repetition penalty	1.1	1.1	1.05

Lower temperature is strongly recommended for competition mathematics and formal reasoning tasks. The model's <think> blocks will be visible in output by default — this is expected behavior and reflects genuine intermediate reasoning.

Training Details

Hardware

Component	Specification
GPU	NVIDIA A100 40GB
Precision	BF16 native
Framework	Unsloth 2026.4 + TRL
Flash Attention	Xformers fallback
Cloud Compute	Google Colab Pro
Training Time	~90 minutes
Total Cost	< $15 USD

LoRA Configuration

Parameter	Value
Rank (r)	16
Alpha	32
Dropout	0.0
Target Modules	q, k, v, o, gate, up, down
Trainable Parameters	18.4M (1.18%)
Gradient Checkpointing	Unsloth smart offload
Quantization	4-bit NF4 (QLoRA)

Optimizer & Training Configuration

Parameter	Value
Optimizer	AdamW 8-bit
Learning Rate	2e-4
LR Scheduler	Cosine
Warmup Steps	100
Weight Decay	0.01
Per-device Batch Size	4
Gradient Accumulation	8
Effective Batch Size	32
Max Sequence Length	4,096 tokens
Epochs	2
Total Steps	~1,376

Dataset

Training used OpceanAI/Yuuki-Personality-v2, a 22,000-example bilingual dataset in ChatML format with native <think> reasoning blocks. The dataset was constructed through a multi-source distillation process:

Kimi K2 — base dataset generation at scale
Gemini — think block generation and reasoning structure
Claude Opus — think block refinement and quality improvement

The dataset covers conversational reasoning, factual Q&A, mathematical problem-solving, code assistance, identity anchoring, and adversarial resistance across English and Spanish.

The RxG Nano fine-tuning objective was identity installation — establishing the YuuKi character over the VibeThinker base without degrading the base model's reasoning capability. This was verified post-training by comparing AIME 2024 scores before and after fine-tuning.

Training Rationale

The choice of VibeThinker-1.5B as base model over alternatives (DeepSeek-R1-Distill-1.5B, Qwen3.5-2B) was informed by benchmark comparison:

Model	AIME 2024	MMLU-Pro	Notes
DeepSeek-R1-Distill-1.5B	28.9%	—	SFT only, no RL stage
Qwen3.5-2B	—	55.3%	Thinking disabled by default at small scale
VibeThinker-1.5B	~80%	~65%	SFT + RL distillation from frontier models

VibeThinker applies both SFT and RL distillation from multiple frontier teachers — the same principle as DeepSeek-R1 distillation, but with a broader and more diverse teacher set. This produces a significantly stronger reasoning foundation at 1.5B scale.

Limitations

Context length. Fine-tuning was conducted at 4,096 tokens. The base model supports longer contexts, but performance on tasks requiring context beyond 4,096 tokens has not been formally evaluated.
GPQA Diamond gap. RxG Nano scores 50.9% on GPQA Diamond, below frontier models (Gemini-2.5-Flash at 82.8%, o3-mini at 76.8%). This benchmark requires graduate-level physics, chemistry, and biology knowledge that is underrepresented in the Yuuki training dataset.
OlympiadBench ceiling. 44.6% reflects the upper bound of competition mathematics capability at 1.5B scale with current training methodology. This is a target for improvement in RxG NxG.
Think block quality. Some <think> blocks inherit boilerplate patterns from the training dataset. Reasoning quality is variable — stronger for mathematics and logic, weaker for open-ended knowledge retrieval.
Safety alignment has not been formally evaluated under adversarial conditions. Not recommended for safety-critical deployment without additional review.
HLE at 8.0%. Humanity's Last Exam performance reflects genuine capability limits at this scale. The score was evaluated using Claude Sonnet 4.6 as judge, which may introduce evaluation variance.

The RxG Family

RxG is the reasoning-specialized lineage within the OpceanAI ecosystem. Each release targets a specific parameter regime and deployment context.

Model	Parameters	Status	Base	Primary Target
YuuKi RxG Nano	1.5B	Released	VibeThinker-1.5B	Edge deployment, reasoning baseline
YuuKi RxG 8B	8B	Released	DeepSeek-R1-Distill-Qwen-8B	General reasoning, competition math
YuuKi RxG VL 27B	27B	Planned	TBD	Multimodal reasoning, flagship

OpceanAI Ecosystem

Model	Family	Parameters	Description
YuuKi RxG Nano	RxG	1.5B	Edge reasoning, AIME 80.0%, TruthfulQA 89.6%
YuuKi RxG 8B	RxG	8B	Reasoning flagship, TruthfulQA 96.6%
Yumo Nano	Yumo	1.5B	Math specialist, surpasses DeepScaleR
YuuKi NxG VL	NxG	7B	General conversation + vision

Links

Citation

bibtex

@misc{awa_omg_2026_rxg_nano,
    author       = { awa_omg },
    title        = { Yuuki-RxG-nano (Revision 1.0) },
    year         = 2026,
    url          = { https://huggingface.co/OpceanAI/Yuuki-RxG-nano },
    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.

Inherits license terms from VibeThinker-1.5B and Qwen2.5-Math-1.5B.

Updates

Date	Milestone
2026-04-27	MMLU-Pro 65.63% — exceeds DeepSeek V3 671B
2026-04-27	AIME 2024 80.0% — 2.77× DeepSeek-R1-Distill-1.5B
2026-04-27	TruthfulQA MC1 89.6% (1-shot) verified
2026-04-27	HLE 8.0% evaluated with Claude Sonnet 4.6 judge
2026-04-27	YuuKi RxG Nano v1.0 released on Hugging Face

Last updated: 2026-04-27

1.5B parameters. 90 minutes of training. Under $15 of compute.
AIME 2024 at 80.0%. MMLU-Pro exceeding a 671B model.
This is what frontier distillation makes possible at the edge.

The RxG family. Built under constraints. No excuses.