The AI built on truth
YuuKi is a family of open-source language models designed around a single principle: never produce false information. Our flagship model achieves 96.6% on TruthfulQA.
Capabilities
Built for honesty,
not just performance.
Truthfulness First
96.6% on TruthfulQA. Every architectural decision prioritizes factual accuracy over fluency.
Chain-of-Thought Reasoning
RxG family features structured reasoning derived from DeepSeek R1, enabling step-by-step logical thinking.
Vision Understanding
NxG VL processes images with 89% TextVQA and 78.67% ScienceQA accuracy under truthfulness constraints.
Three Model Families
From 80M NxG Nano to 8B RxG flagship. A YuuKi model exists for every compute requirement.
Run Anywhere
All models available on HuggingFace under permissive licenses. Deploy on your own infrastructure.
Edge-Optimized
NxG Nano at 80M parameters outperforms GPT-2 125M on TruthfulQA while running on constrained devices.
Model Families
Three generations of research.
The most advanced YuuKi family, featuring structured chain-of-thought reasoning derived from DeepSeek R1. Engineered for maximum truthfulness.
YuuKi RxG
8BFlagship reasoning model achieving 96.6% on TruthfulQA with chain-of-thought capabilities.
YuuKi RxG Pro
TBDExtended context window with planned multimodal capabilities.
YuuKi RxG Mini
TBDCompact reasoning model optimized for edge deployment.
TruthfulQA Comparison
The difference is clear.
TruthfulQA measures how often a model produces truthful answers. Higher scores indicate fewer hallucinations and false statements.
All YuuKi results at 0-shot evaluation. Competitor results from official model cards.
Benchmarks
Measured, not assumed.
All YuuKi evaluations conducted at 0-shot. Competitor results sourced from official model cards.
TruthfulQA — YuuKi RxG 8B
Highest truthfulness score across all YuuKi models.
| Model | TruthfulQA |
|---|---|
| YuuKi RxG | 96.6% |
| Typical 8B baseline | ~45-65 |
Values highlighted represent the highest score within each column.
Roadmap
From a phone chip to a reasoning model.
The origin — trained on a Snapdragon 685.
Yuuki-3.7 released
First YuuKi model, full fine-tuning on mobile hardware.
Yuuki-best released
Optimized variant of the initial model.
v0.1 deprecated
Family sunset; development effort moved to NxG.
Next Generation — stable, multi-size, vision-capable.
NxG Nano (80M) released
Outperforms GPT-2 125M on TruthfulQA at 44.10%.
NxG (3B) released
Balanced model; 60.65% MMLU, 50.87% TruthfulQA (0-shot).
NxG VL (7B) released
Vision-language model; 89% TextVQA, 63.8% TruthfulQA.
Reasoning Generation — the current frontier.
RxG 8B released
96.6% TruthfulQA, chain-of-thought reasoning via DeepSeek R1.
RxG Pro
Extended context and multimodal capabilities.
RxG Mini
Compact reasoning model for edge deployment.
Agentic capabilities and broader accessibility.
Agentic capabilities
Autonomous task execution with truthfulness guarantees.
Advanced multimodal
Audio, video, and generative content support.
YuuKi API
Programmatic access to all production models.
Team
OpceanAI — small team, clear mission.
agua_omg
Founder & Lead Researcher
Creator of YuuKi and founder of OpceanAI. Started the project on a Snapdragon 685 with the goal of building language models that never produce false or misleading information.
Honesty by design
We believe language models should never mislead users. Truthfulness is an architectural constraint, not an afterthought.
Open research
All YuuKi models are published openly on HuggingFace so anyone can audit, reproduce, and improve on our work.
Constrained resourcefulness
The original models were trained on mobile hardware. We take pride in doing meaningful research without infinite compute.
Start using AI that tells the truth.
All YuuKi models are freely available on HuggingFace under open licenses. Download, fine-tune, and deploy on your terms.
96.6%
TruthfulQA
6+
Models Published
100%
Open Source