sayeed105236

CuttyMOA-1.3

README

License: apache-2.0

Model details

Table

Base architecture	Qwen3.8-27B (dense, hybrid linear/full attention)
Parameters	27B (merged bf16, ~51 GB)
Fine-tuning method	LoRA r=64, α=128, dropout 0.05
Target modules	attention (q/k/v/o) · linear-attention DeltaNet (in_proj_qkv/in_proj_z/out_proj) · MLP (gate/up/down)
Precision	bf16, gradient checkpointing, Triton flash-linear-attention kernels
Context window	262,144 tokens native
Max output	8,192 tokens
Learning rate	1.5e-4 cosine, warmup 200
Effective batch	16 rows / optimizer step (bs 2 × accum 8)
Steps	600 optimizer steps, 1 epoch
Framework	transformers + peft + accelerate
Hardware	NVIDIA RTX PRO 6000 Blackwell 96 GB (vast.ai)
Training time	2 h 18 m (13.5 s/step)
Loss	5.33 → 0.73 (logged every 5 steps)

Capabilities & skills

CuttyMOA-1.3 is the reasoning core of the CUTEADMOA platform. Inherited strengths from the base architecture, amplified by domain fine-tuning:

Chat & reasoning — long-horizon multi-step tasks, instruction following, summarisation
Code — generation, review, debugging, SQL (code is the largest corpus category)
Security intelligence — 17 cybersecurity intents with severity classification, feeding the platform's HTML/DOCX report engine
Research — 24 intent categories, multilingual (EN/BN/HI/AR/ZH/JA/KO/FR/DE/ES voices and data paths across the platform)
Agentic tool-calling — trained on function-calling corpora (BFCL, xLAM, Glaive, Hermes)
Scraping-adjacent data tasks — extraction planning and JSON/CSV schema design for the ScrapePower engine (Selenium + AI extraction)
262K long context — whole codebases, contracts and books in one prompt
Vision-language — native image/video understanding inherited from the Qwen3.8 base (fine-tune kept the vision tower; text performance is the training focus)

Platform integration (what ships with it)

Table with columns: Component, Detail
Component	Detail
Unified MOA router	23+ engines, smart routing, cross-model aggregation
Playground	cuteadmoa.site — live CuttyMOA-1.3 demo, 4 modes (fast/auto/medium/thinking)
OpenAI-compatible API	`/v1/chat/completions`, Bearer key auth
ScrapePower	enterprise scraping + AI extraction, cron scheduling
Native apps	6 macOS apps + iOS/iPad companion
Voice	STT + TTS across 3 engines (Samantha default)
Continuous training	204-verified dataset loop + live agent-pair harvesting

Training data

Corpus

The CUTEADMOA corpus combines:

204-verified domain datasets (live-checked on Hugging Face, 4.49M pairs total) — chat, code, SQL, agentic tool-calling, cybersecurity, multilingual, reasoning/math, finance, medical, legal, science, sentiment, RAG, enterprise documents. A 40,652-row stratified sample was used for this run.
Live-harvested agent pairs — 2,337 curated instruction-response pairs extracted from real production agent sessions (chat, code, security, ecommerce, automation, ops, research, scraping, design, writing), cleaned of session scaffolding, oversampled ×4 to ~19% of the corpus and interleaved for uniform exposure.

Identity injection: every training row carries the system identity (CuttyMOA-1.3 · Developed by MD ABU SAYEED | Powered by DOTPROGRAMMERS | Platform: CUTEADMOA), so the model self-identifies correctly.

Training procedure

Single-epoch bf16 LoRA SFT with the CUTEADMOA identity prompt on every row; cosine LR, warmup 200, grad-norm 1.0, sequence packing at 1,024 tokens. Loss curve above is the real logged series (500 points, checkpoint-500 trainer_state.json). Checkpoints 250/500 with optimiser state are archived in CuttyMOA-1.3-checkpoints (private backup repo).

Evaluation & speed (measured, not estimated)

Speed

Table with columns: Deployment, Tokens/s
Deployment	Tokens/s
Public demo (lab CPU, llama.cpp Q4_K_M)	1.8–2.0
GPU pod serving (bf16, RTX-class)	25–45 (measured under train-adjacent load)
Small sibling 4B (local M4 est.)	25–55

The 4B sibling CuttyMOA-1.3-Small is trained on the identical corpus and identity, built for local/Ollama use: sayeed105236/CuttyMOA-1.3-Small / -Small-GGUF.

Version history

Table with columns: Version, Date, Base, Focus
Version	Date	Base	Focus
CuttyMOA-1.0	2026-08	dense 8B	first full pipeline — 120→204 datasets proven
CuttyMOA-1.2	2026-08	Qwen3.6-35B-A3B MoE	MoE scale-up, 204 verified datasets, 24 categories
CuttyMOA-1.3	2026-09	Qwen3.8-27B	flagship: hybrid linear-attention arch, live-agent data, this model
1.3-Small	2026-09	Qwen3-4B

Usage

bash
# OpenAI-compatible API
curl https://cuteadmoa.site/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"CuttyMOA-1.3",
       "messages":[{"role":"user","content":"Hello!"}],
       "max_tokens":512}'

python
from openai import OpenAI
client = OpenAI(base_url="https://cuteadmoa.site/v1", api_key="YOUR_KEY")
r = client.chat.completions.create(model="CuttyMOA-1.3",
    messages=[{"role":"user","content":"Explain RAG in two sentences."}])

Serve locally with llama.cpp (Q4_K_M GGUF ~17 GB — needs 20+ GB RAM): llama-server -m CuttyMOA-1.3-Q4_K_M.gguf -c 32768 -ngl 99

Guidelines & limitations

Identity: the model is trained to identify as CuttyMOA-1.3 by MD ABU SAYEED — please don't strip attribution when redistributing fine-tunes of it.
Safety: not safety-evaluated for open-ended consumer use; intended as the CUTEADMOA platform's private inference engine. Guardrails inherited from the base model.
Limits: knowledge cutoff = base model (Qwen3.8); English/Bangladeshi-market emphasis in the live-pair data; may occasionally leak the identity block on identity-probing prompts — this is by design.
Hallucination: like all LLMs, verify factual/code output; spot-check categories ship with the platform's evaluation loop.

Citation & links

bibtex
@misc{cuttymoa13,
  author = {MD ABU SAYEED},
  title  = {CuttyMOA-1.3: Self-Trained Mixture-of-Agents Flagship (CUTEADMOA)},
  year   = {2026},
  url    = {https://huggingface.co/sayeed105236/CuttyMOA-1.3},
  note   = {Powered by DOTPROGRAMMERS}
}

🔗 LinkedIn: https://www.linkedin.com/in/abu-sayeed-6296a1b9
🌐 Platform + live playground: https://cuteadmoa.site
📦 GGUF build: https://huggingface.co/sayeed105236/CuttyMOA-1.3-GGUF
🧩 Small sibling: https://huggingface.co/sayeed105236/CuttyMOA-1.3-Small
🗂️ Previous: CuttyMOA-1.0 · CuttyMOA-1.2

License

Apache-2.0 (inherited from the base model).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

sayeed105236

Model Tree

Base

Qwen/Qwen3.8-27B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality