CUTEADMOA — Self-Trained AI Models & Unified Agent Platform

Q: Is the API OpenAI-compatible?

Yes. Point any OpenAI-compatible client at the base URL https://cuteadmoa.site/v1 with a bearer key and use the standard chat completions payload.

Q: Does it call itself CuttyMOA-1.3?

Yes. The identity is injected into every training row, so the model self-identifies as CuttyMOA-1.3, developed by MD ABU SAYEED on the DOTPROGRAMMERS CUTEADMOA platform.

Q: Is the training data public?

The domain datasets are verified public corpora. The live agent pairs are derived from the lab's own production sessions and are cleaned of private content before use.

262K context204 verified datasetsLive agent pairsOpenAI-compatible API17 security intentsScrapePower engineSelenium automationGGUF quantisationVoice STT & TTS6 macOS apps + iOSLoRA fine-tuningContinuous learning loop 262K context204 verified datasetsLive agent pairsOpenAI-compatible API17 security intentsScrapePower engineSelenium automationGGUF quantisationVoice STT & TTS6 macOS apps + iOSLoRA fine-tuningContinuous learning loop

Live Demo

Playground — chat with CuttyMOA-1.3

No signup. Rate-limited public demo served straight from the lab's own hardware. Full-speed API keys are available for the OpenAI-compatible endpoint.

cuteadmoa.site — CuttyMOA-1.3 playground LIVE

👋 Hello! I'm CuttyMOA-1.3 — ask me anything: reasoning, code, writing, multilingual chat. I run on the CUTEADMOA lab server.

Platform

Capabilities & skills

Everything below runs on the same self-hosted stack, orchestrated around the CuttyMOA model family.

💬

Unified Chat & Aggregation

23+ engines behind one smart router, with cross-validated multi-engine answers for higher reliability.

Smart Router

🧠

CuttyMOA Model Family

Self-trained models across four generations — 1.0, 1.2 MoE, the 1.3 flagship, and a compact 1.3-Small.

Self-trained

⌨️

Code & SQL

Generation, review and debugging across languages — code is the single largest slice of the training corpus.

Largest corpus slice

🛡️

Security Intelligence

17 cybersecurity intents with severity classification, feeding HTML and DOCX report generation.

17 intents

🕷️

ScrapePower Engine

Selenium Chrome automation, 8 pre-built extractors, AI-powered extraction, JSON/CSV export and cron scheduling.

Enterprise scraping

📚

262K Long Context

Entire codebases, contracts and books inside a single prompt window — natively 262,144 tokens.

Long context

🤖

Agentic Tool-Calling

Trained on function-calling corpora including BFCL, xLAM, Glaive and Hermes tool-use sets.

Function calling

🔊

Voice — STT & TTS

Speech-to-text and text-to-speech across three engines, wired into native app voice flows.

3 engines

🖥️

Native Apps

Six macOS control applications plus an iOS/iPad companion for full desktop and mobile integration.

6 apps + iOS

How it works

From raw data to a served model

A transparent, reproducible pipeline — every stage runs on infrastructure this lab controls.

1 · Corpus assembly

Two data streams are merged into one training corpus: a stratified sample of the 204-verified domain datasets, plus live instruction-response pairs harvested from real production agent sessions.

Session scaffolding stripped from harvested pairs
Degenerate rows discarded before training
Local pairs oversampled and interleaved for even exposure

Organised blue network cables representing structured data flow — Corpus assembly — two streams, one clean dataset

2 · LoRA fine-tuning on GPU

The flagship runs bf16 LoRA training on a 96 GB GPU host: rank-64 adapters across attention, linear-attention and MLP projections, with gradient checkpointing and compiled kernels.

600 optimizer steps over a 50,000-row corpus
Loss fell from 5.33 to 0.73
Checkpoints archived and auto-backed up during the run

Glowing server racks in a high-tech data centre — GPU fine-tuning — 600 steps, checkpointed throughout

3 · Merge, quantise, serve

Adapters are merged back into the base model, then quantised to GGUF for efficient serving. The same weights ship in three forms: full merged model, GGUF build, and a compact 4B sibling.

Merged bf16 model published to Hugging Face
GGUF Q4_K_M build for llama.cpp and Ollama
4B sibling for local inference on a laptop

Glasses reflecting lines of code — Merge, quantise, serve — one set of weights, three formats

Benchmarks

Measured, not claimed

Real logged training metrics and real generation timings — pulled from the actual training run and live endpoint tests.

Training loss curve falling from 5.33 to 0.73 over 500 logged steps

Training loss — 5.33 → 0.73500 logged points, checkpoint trainer state

Donut chart showing corpus composition by source

Corpus composition50,000 rows: domain datasets + live agent pairs

Bar chart of measured generation speed by deployment target

Generation speed by deploymentMeasured on the public endpoint and GPU pod


Deployment target	tokens / second	Notes
Public demo — lab CPU, GGUF Q4_K_M	1.8 – 2.0	Rate-limited playground endpoint
GPU pod — bf16 serving	25 – 45	On-demand burst capacity
Small 4B — local laptop	25 – 55	Estimated for Apple silicon Q4_K_M
Flagship 27B on 16 GB laptop	cannot load	Needs ~17 GB even at Q4

Integrate anywhere

OpenAI-compatible API access

Point any OpenAI-compatible client at the lab endpoint. Model identifier, limits and a working request below.

Endpoint

Chat Completions

Base URLhttps://cuteadmoa.site/v1

Chat route/v1/chat/completions

Model IDCuttyMOA-1.3

AuthBearer <key>

Limits

Context & output

Context length262,144 tokens

Max output8,192 tokens

Modesfast · auto · medium · thinking

FormatOpenAI chat completions

Quick call

One request

# use your lab key
curl https://cuteadmoa.site/v1/chat/completions \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "CuttyMOA-1.3",
    "messages": [{"role":"user",
      "content":"Hello!"}],
    "max_tokens": 512
  }'

Model family

Four generations, trained in the open

Each version is checkpointed and published — nothing is lost between runs.

CuttyMOA-1.3 — the flagship

A 27-billion-parameter model with a native 262,144-token context window. Trained with bf16 LoRA over the full CUTEADMOA corpus, then merged and quantised for serving.

600 optimizer steps · loss 5.33 → 0.73
Hybrid linear-attention architecture
Merged bf16 model published, ~51 GB

Friendly white robot assistant representing the CuttyMOA model — CuttyMOA-1.3 — 27B, 262K context

CuttyMOA-1.3-Small — built for your laptop

The 27B flagship needs at least 17 GB even when heavily quantised, so it cannot run on a 16 GB machine. The Small sibling is trained on the identical corpus and identity, on a compact 4B base — about 2.5 GB at Q4_K_M.

Same corpus, same identity injection
Runs locally in Ollama at interactive speed
Published alongside the flagship

Developer working across multiple monitors — CuttyMOA-1.3-Small — local inference that actually fits

Generation	Base	Focus	Status
CuttyMOA-1.0	Dense 8B	First full training pipeline	Published
CuttyMOA-1.2	35B-A3B MoE	Mixture-of-experts scale-up	Published
CuttyMOA-1.3	Qwen3.8-27B	Flagship — 262K context, hybrid attention	Live
CuttyMOA-1.3-Small	Qwen3-4B	Local and edge inference	In pipeline

Operator monitoring multiple screens in a control room — Continuous monitoring and evaluation loop

Security & operations

The platform ships with a cybersecurity intent layer and an ongoing evaluation loop that keeps the models honest in production.

17 security intents with severity classification
Generated HTML and DOCX security reports
Token-based auth with scheduled rotation
Continuous corpus harvesting and validation

Roadmap

What comes next

Concrete, measured steps — not aspirational slideware.

NOW

Publish the Small sibling and GGUF builds

CuttyMOA-1.3-Small plus quantised GGUF packages for llama.cpp and Ollama, so the model runs on ordinary laptops.

Faster training, deeper coverage

Sequence packing and premium GPU hosts to roughly double throughput — enough to see every live agent pair instead of a subset.

THEN

Scale parameters with mixture-of-experts

A 35B-A3B mixture-of-experts variant: more total parameters while staying fast, because only a fraction activates per token.

ONGOING

Evaluation and release cadence

Identity retention, scaffolding leakage and per-category spot checks, feeding a repeatable release loop.

White paper

The full architecture, in print and on film

A 26-page technical white paper covering the routed model portfolio, the credential fabric, the proxy fleet, governed agent execution and the self-training loop — with a narrated video walkthrough of the same material.

26pages, typeset

11narrated chapters

52timed captions

Read the white paper Download PDF

Title card of the CUTEADMOA white paper walkthrough video

▶

Peer-reviewed publication

The platform, published in the literature

Advances in Artificial Intelligence and Robotics Research · Vol. 2 · No. 3 · pp. 190–208 · September 2026

CUTEADMOA: A Unified Mixture-of-Agents Platform with Continuous Self-Training for Autonomous Enterprise AI Automation

Md Abu Sayeed · Scientific Research Publishing · Published 24 September 2026
DOI 10.4236/airr.2026.23011 · ISSN 3143-3987 (print) / 3143-3995 (online)

Read the paper Full-text PDF DOI

Cite this work

@article{sayeed2026cuteadmoa,
  author  = {Md Abu Sayeed},
  title   = {CUTEADMOA: A Unified Mixture-of-Agents Platform with Continuous Self-Training for Autonomous Enterprise AI Automation},
  journal = {Advances in Artificial Intelligence and Robotics Research},
  volume  = {2},
  number  = {3},
  pages   = {190--208},
  year    = {2026},
  issn    = {3143-3987},
  doi     = {10.4236/airr.2026.23011},
  url     = {https://www.scirp.org/journal/paperinformation?paperid=154162}
}

FAQ

Questions, answered

What exactly is CuttyMOA-1.3?

It is a 27-billion-parameter language model self-trained by CUTEADMOA LAB on 204 verified datasets plus live agent sessions. It supports a 262,144-token context window and is served through an OpenAI-compatible API.

Can I run it on my own laptop?

The flagship needs at least 17 GB of memory at Q4 quantisation, so it will not load on a 16 GB machine. For local use we ship CuttyMOA-1.3-Small, a 4B sibling trained on the same corpus, which runs comfortably in Ollama.

Is the API OpenAI-compatible?

Yes. Point any OpenAI-compatible client at the base URL with the model identifier CuttyMOA-1.3 and a bearer key, and use the standard chat completions payload.

How long does a response take?

On the public CPU demo, expect roughly 15 to 60 seconds depending on reply length. On the GPU burst service the same model runs far faster. Modes let you trade depth for latency.

Does it call itself CuttyMOA-1.3?

Yes — the identity is injected into every training row, so the model self-identifies as CuttyMOA-1.3, developed by MD ABU SAYEED, on the DOTPROGRAMMERS CUTEADMOA platform.

Is the training data public?

The domain datasets are verified public Hugging Face corpora. The live agent pairs are derived from this lab's own production sessions and are cleaned of private content before use.

Work with us

Build and train your own AI model

Bring your data. We shape it, train a model that speaks your domain, and put it behind an API or chat interface you own — no research team required.

From your data to a working model

A transparent pipeline, run with you at every milestone. You approve each stage before we move on — and you keep ownership of the data and the model.

Scope call — define the task and what "good" looks like
Data preparation — clean, structure and format your corpus
Training — fine-tune or build on your domain, with logged metrics
Deployment — ship behind an API or chat interface you control

Request a fit assessment See a live model

Developer building and training a custom AI model — Training runs — scoped, checkpointed, reproducible

🧹

Data readiness

We audit your corpus, flag what can't train, and tell you what to fix before a single GPU hour is spent.

Step 01

🧱

Corpus preparation

Cleaning, deduplication, structuring and formatting into a dataset that trains reliably.

Step 02

🎛️

Custom training

LoRA fine-tuning or a full build on your domain, with logged loss and evaluation checkpoints.

Step 03

🧩

Merge & quantise

Adapters merged back into the base and packaged in the formats you actually deploy, including GGUF.

Step 04

🚀

Deploy & serve

Behind an OpenAI-compatible API or a chat interface you control, with monitoring from day one.

Step 05

🔁

Ongoing evaluation

Quality and identity checks feeding a repeatable update loop, so the model keeps improving.

Step 06

Early access

Tell us what you want to build

Share a few details and we'll come back with a scoped plan and next steps. We only contact people who opt in here.

Why teams work with this lab

You get a model tuned to your material, not a generic endpoint you have to re-explain every prompt.

Your data stays yours — documented, exportable, never resold
Fixed milestones instead of open-ended hourly work
Runs on the same stack behind the live playground above
Reply within 1–2 working days

Prefer a quick chat? Add your WhatsApp or Telegram in the form and that's the fastest way to reach you.

Or email us directly at hello@dotaistudio.cloud — every enquiry gets an automatic acknowledgement.

Full name *

Work email *

WhatsApp (with country code)

Telegram username

What do you want to do? *

Country

Company / organisation

What should the model do?

Inside the lab

Infrastructure that trains and serves real models

Training hardware, serving infrastructure and the tooling around it — the same stack behind the live playground above.

Server rack hardware powering model training and serving — Serving infrastructure

Robotic arm representing applied AI and automation — Applied AI & automation

Abstract blue light network representing data flow — Data flow at scale

Playground — chat with CuttyMOA-1.3

Capabilities & skills

Unified Chat & Aggregation

CuttyMOA Model Family

Code & SQL

Security Intelligence

ScrapePower Engine

262K Long Context

Agentic Tool-Calling

Voice — STT & TTS

Native Apps

From raw data to a served model

1 · Corpus assembly

2 · LoRA fine-tuning on GPU

3 · Merge, quantise, serve

On-demand GPU burst

Measured, not claimed

OpenAI-compatible API access

Chat Completions

Context & output

One request

Four generations, trained in the open

CuttyMOA-1.3 — the flagship

CuttyMOA-1.3-Small — built for your laptop

Security & operations

What comes next

Publish the Small sibling and GGUF builds

Faster training, deeper coverage

Scale parameters with mixture-of-experts

Evaluation and release cadence

The full architecture, in print and on film

The platform, published in the literature

CUTEADMOA: A Unified Mixture-of-Agents Platform with Continuous Self-Training for Autonomous Enterprise AI Automation

Cite this work

Questions, answered

Build and train your own AI model

From your data to a working model

Data readiness

Corpus preparation

Custom training

Merge & quantise

Deploy & serve

Ongoing evaluation

Tell us what you want to build

Why teams work with this lab

Request received.

Infrastructure that trains and serves real models

From raw data to a served model

Start building on CuttyMOA-1.3