
👤 Developed by	MD ABU SAYEED
⚡ Powered by	DOTPROGRAMMERS
🚀 Platform	CUTEADMOA (Mixture-of-Agents)
🧠 Base Architecture	Qwen 3.6 35B A3B (MoE)
📦 Format	GGUF (Q4_K_S — CPU-optimized)
🎯 Active Params	~3B per token (35B total, 128 experts)
📊 Context Window	8,192 tokens (extensible to 32K)
🔒 License	Apache 2.0

Overview

CuttyMOA-1.2 is a CPU-first Mixture-of-Experts language model built on the Qwen 3.6 35B A3B architecture. With only ~3B parameters active per token from a 35B expert pool, it achieves strong reasoning performance while running efficiently on CPU hardware — no GPU required.

It is the direct successor to CuttyMOA-1.0, upgraded from a 15B dense model to 35B MoE with 2.3× more total knowledge capacity.

Why CuttyMOA-1.2?

Feature	CuttyMOA-1.0	CuttyMOA-1.2
Architecture	Dense 15B	MoE 35B (3B active)
Context	8K	8K (extendable to 32K)
GGUF Size	8.4 GB (Q4_K_M)	17 GB (Q4_K_S)
Inference Speed (CPU)	8-12 t/s	2-4 t/s
Knowledge Breadth	Moderate	2.3× larger expert pool
Training Data	Base only	Base + 857K curated pairs

Quick Start

Option 1: llama.cpp (Recommended for VPS)

bash

# Download the GGUF
huggingface-cli download sayeed105236/CuttyMOA-1.2 \
  CuttyMOA-1.2-Q4_K_S.gguf --local-dir ./

# Run with llama-server
./llama-server \
  -m CuttyMOA-1.2-Q4_K_S.gguf \
  -c 8192 -n 1024 \
  --port 5401 --host 127.0.0.1

# Test
curl http://127.0.0.1:5401/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Who are you?"}],"max_tokens":50}'

Option 2: Ollama

bash

# Download GGUF + Modelfile from this repo, then:
ollama create cuttymoa-1.2 -f Modelfile
ollama run cuttymoa-1.2

Option 3: CUTEADMOA Integration

CuttyMOA-1.2 plugs directly into the CUTEADMOA Mixture-of-Agents platform as a local inference engine. See cuteadmoa.site for platform documentation.

Hardware Requirements

Setup	RAM	Disk	Notes
Minimum	22 GB	18 GB	Q4_K_S, 4K context
Recommended	28 GB	20 GB	Q4_K_S, 8K context
Optimal	32 GB	25 GB	Q4_K_M, 8K context

Runs on any x86_64 CPU. Tested on AMD EPYC, Intel Xeon, Apple M-series.

Training Data

The model ships with the CUTEADMOA Knowledge Corpus:

857,223 instruction-response pairs across 24 domains
Domains: code, security, news, medical, legal, finance, science, multilingual, RAG, agentic-tools, reasoning, dialogue, and more
Full dataset available at sayeed105236/cuttymoa-training-corpus

Future Fine-Tuning

QLoRA fine-tuning recipes will be published for users who want to adapt CuttyMOA-1.2 to their own domains. Budget ~$2-4 on RunPod A40 for a full fine-tuning run.

CUTEADMOA Ecosystem

CuttyMOA-1.2 is the local inference engine for CUTEADMOA, a Mixture-of-Agents platform combining:

23+ text models across 6 providers
24 security intents with HTML/DOCX reports
Enterprise web scraping (ScrapePower)
Native macOS & iOS apps
Smart routing with multi-model aggregation

Version History

Version	Date	Base	Params	Notes
CuttyMOA-1.0	2026-08	Custom	15B dense	First release
CuttyMOA-1.2	2026-08	Qwen 3.6	35B MoE	Current