OpceanAI
Born on a Snapdragon 685 phone. Built through constraints, not resources. Now an ecosystem of models, systems, and infrastructure.
Not one project. A growing system of ideas that learned how to become real.
A technology organization focused on the development of artificial intelligence models, infrastructure, and systems.
It is not a single product. It is an evolving ecosystem of software, research, architecture, and applied experimentation.
OpceanAI is the result of years of curiosity, resource limitations, hands-on experimentation, and long-form technical iteration. Its evolution reflects a path from simple bots to foundational AI work, then to system-level software and infrastructure.
“OpceanAI exists to build systems that feel intentional, useful, and technically alive.”
Where it began
OpceanAI was born on April 23, 2023, originally under the name Ocean. At that stage, the mission was simple: build bots for Discord and Telegram. That period became known as Bots New.
Sakura
- Written as a single
main.py - Around 11k lines of code
- A massive monolithic beginning
- A direct, bold, experimental approach
Nebula
- Built with JS/TS
- A different implementation style
- A contrasting path to Sakura
- Part of the same foundational era
The DNA of this era
“Sakura and Nebula were not just the first projects. They were the first signs of a system becoming an organization.”
The first encounter with LLMs
While the creator of OpceanAI, awa-omg, was thinking about how to build a first model of AI for Sakura, the first interaction with LLMs happened.
At that moment, there was no deep familiarity with the modern AI ecosystem.
The environment was still very early
- No deep knowledge of Hugging Face
- No deep familiarity with modern AI tooling
- No “instant maturity” in the stack
- Just curiosity, code, and a growing ambition
This moment represents the transition from “building bots” into “building intelligence.”
It is the conceptual bridge between simple automation and model-based systems.
OpceanAI Lab
In 2025, awa decided to create something historic: YuuKi v0.1. This became the first prototype of the OpceanAI Lab era.
The origin of the name YuuKi
The name YuuKi comes from a very specific personal and emotional context.
Around October 2025, after personal problems, awa entered a period of depression. During that time, Girls' Last Tour became deeply meaningful. Its style, atmosphere, and emotional tone made a strong impact.
After finishing the manga, awa felt unable to continue carrying that emotional state in the same way. A Discord bot called Yuki was created, but it did not become what was hoped for, and it was eventually discontinued.
Even so, that emotional path remained important, because it became part of the name and identity of what would later become YuuKi.
Why this matters
This part of the story should not be treated like trivia. It is part of the emotional origin of the lab era.
The project did not emerge from blank product strategy. It emerged from a real attempt to give form to an internal vision.
The first training attempt
In December 2025, awa had a new idea:
“What if I make an AI?”
So a dataset was downloaded — essentially all of Wikipedia — and the work began.
The tools were very limited
The first attempt used TinyGram. But it became clear that the ambition exceeded the environment, so PyTorch had to be installed.
Environment constraints
At that time, the device available was only a Snapdragon 685 phone. The setup effort was enormous.
This was a difficult period, but it was also foundational. OpceanAI was never built with comfortable resources. It was built through persistence, adaptation, repeated technical struggle, and learning through friction.
Iris
As January arrived, the technical learning process continued. The first AI was born on January 12, 2026, under the codename:
Iris
But that name was not the final identity.
The final name became YuuKi, chosen as a tribute and transformation: a reference to Yuu from Girls' Last Tour, combined with a Japanese snow-like suffix, transformed into a name that felt more fitting for the project.
Why the rename mattered
This rename turned a technical milestone into an identity.
Iris was the first milestone. YuuKi became the real symbol.
2.66 years
The training time estimate was far too long: about 2.66 years on average.
That made the approach feel impossible.
Awa wanted to stop. Wanted to do nothing. Wanted a better path.
The architectural discovery
Then came a video about BitNet and the idea of full fine-tuning — retraining all the weights of a model.
This inspired a new direction.
GPT-2 82M
LLaMA 3.2 1B was downloaded, but the phone could not realistically handle that scale.
So the next attempt was much smaller: GPT-2 82M. That became the practical breakthrough point.
This moment represents the move from impossible ambition to constrained but real execution.
This should feel like the “hard part” of the keynote: quiet, serious, and impressive without overexplaining.
From YuuKi v0.1 to YuuKi RxG
YuuKi is the central intellectual line of OpceanAI. It is not just a model name. It is a lineage.
v0.1
First prototype
The first prototype of the OpceanAI Lab era. Where it all began.
NxG
Next generation
The continuing development of the original AI line.
RxG
Refined generation
The current state of the YuuKi lineage.
Beyond YuuKi
OpceanAI does not only create general-purpose systems. It creates targeted, domain-oriented models and explores multiple paths toward intelligence.
Yumo
The Yumo models emerged as a specialized branch of the YuuKi ecosystem. Based on YuuKi, but specialized in mathematics. Demonstrates that OpceanAI creates targeted, domain-oriented models — specialized, focused, mathematical, structured.
OwO
Short, memorable, and identity-driven. Shows that OpceanAI uses naming not only as branding, but as architectural identity.
OvO
Origin and versioning as architectural identity. Paired with OwO — one for reasoning, one for origin/versioning.
Yaki
Based on YuuKi, enhanced with multimodal abilities. Capabilities injected via LLaVA. A non-native VL model — different from YuuKi VL models. Shows that OpceanAI is not limited to one modality or one architecture style.
Imprint
The Imprint line expands the ecosystem into multimodal territory alongside Yaki. Exploring multiple paths toward intelligence.
Tsuki
A token compression model created as a quiet contribution to the ecosystem. Trained on 4,160 bilingual examples (Spanish and English), six different task types. Result: average 57.6% token reduction. Its value is precision, not loud branding.
Tsuki — Token Compression
A quiet contribution to the ecosystem
4,160
Bilingual examples
6
Task types
57.6%
Token reduction
Tsuki was trained on Spanish and English examples across six different task types. The result was an average 57.6% token reduction.
“Quality over noise, even without recognition.”
Built through constraints, not resources
0
Founded
0+
Models Created
0
Training Examples
0%
Token Reduction
0
Years of Training Evaded
Doki
Bringing Docker containers to Android.
The most recent project, launched in May 2026. Doki is a system focused on bringing Docker containers to Android.
It is compatible with OCI images and supports four layers of isolation depending on the environment.
OCI Compatible
Full compatibility with OCI container images. Run standard Docker containers on Android.
4 Isolation Layers
Multiple security layers depending on the environment. Adaptable isolation from sandboxed to full system-level.
Android Native
Bringing Docker containers directly to Android devices. No cloud dependency, no remote server.
Doki is important because it represents the move from model research toward platform infrastructure.
This is not only about AI anymore. It is about running systems.
ASL
Android Subsystem for Linux
An alternative to Microsoft's WSL. The next step for OpceanAI.
ASL is focused on implementing a complete Linux kernel running in userspace, so that Linux applications believe they are running on native Linux rather than Android.
How it works
To the Android system, ASL appears as just another process consuming resources.
But internally, it is a real kernel.
Technical direction
C++
Core systems programming and performance-critical paths
Rust
Memory safety and systems-level correctness
Modified proot
Syscall handling layer for Android compatibility
OpceanAI will not build a distribution. Only the kernel. The rest is intended for the community.
ASL development begins once Doki v1 is complete.
NHE — Not Humanity Exam
Every existing benchmark — HLE, MMLU, BIG-Bench, ARC — measures nearly the same dimension:
- What a model knows
- How much human knowledge it can reproduce
- How accurately it can reason
NHE asks a fundamentally different question:
Not how much the model knows.
But how human it still thinks.
NHE measures the presence of six cognitive patterns structurally embedded within human language itself. These are patterns that systems trained on human text cannot fully escape regardless of scale, capability, or intelligence level.
The Imprint Theory
NHE serves as an empirical implementation of The Imprint Theory.
Rather than measuring accumulated knowledge, NHE attempts to measure traces of human cognitive structure remaining inside artificial systems.
The full ecosystem
OpceanAI is not built around a single project. Over time it evolved into an ecosystem of models, systems, experiments, infrastructure, and research.
General Contact
opceanai@gmail.comPersonal Contact
aguitachan3@gmail.comBusiness Contact
contact@opceanai.comOpceanAI is a technology organization that grew from experimentation into a real ecosystem.
It is a story of:
Bots becoming models.
Models becoming systems.
Systems becoming infrastructure.
Infrastructure becoming a research identity.
And that story should be told with the same level of care and precision that the work itself demands.
“OpceanAI is not one project. It is a growing system of ideas that learned how to become real.”