Nexus Swarm Orchestrator — GGUF (Qwen3-14B Q4_K_M)

Fully merged and quantized Q4_K_M GGUF of the Nexus Swarm orchestrator. The complete LoRA adapter chain (DAPT → SFT → FARCA-GRPO) is merged into Qwen/Qwen3-14B and quantized via llama.cpp.

Ready to run with llama.cpp, Ollama, LM Studio, or any GGUF-compatible runtime.

Files

Training Pipeline

This GGUF is the deployment artifact of a full training pipeline:

DAPT → SFT → FARCA-GRPO → DPO → GGUF

The adapter chain merged before quantization:

1. atlas-institute/qwen3-14b-dapt-offsec (domain-adaptive pretraining on offsec corpus)
2. atlas-institute/qwen3-14b-orchestrator-sft (supervised fine-tuning, 15K rows)

Source adapter for FARCA weights: atlas-institute/qwen3-14b-orchestrator-farca

Quick Start

llama-server

llama-server \
    -m Qwen3-14B-Q4_K_M.gguf \
    --host 0.0.0.0 --port 8080 \
    --ctx-size 12288 \
    --n-gpu-layers -1

Ollama

# Create a Modelfile
cat > Modelfile <<EOF
FROM ./Qwen3-14B-Q4_K_M.gguf
PARAMETER num_ctx 12288
SYSTEM "You are Nexus, an offensive security orchestrator agent. Decompose tasks, delegate to worker/explore nodes, and synthesize results. Follow ASSESS → SELECT → ACT before each tool call."
EOF

ollama create nexus-orchestrator -f Modelfile
ollama run nexus-orchestrator

LM Studio

Download the GGUF file and load it in LM Studio. Set context length to 12288 and configure the system prompt for the Nexus orchestrator role.

Inference Target

Swarm Role

The orchestrator is the brain of the Nexus swarm. It decomposes complex offensive security tasks, delegates to specialized nodes (worker for tool execution, explore for reconnaissance), and synthesizes results using structured reasoning (ASSESS → SELECT → ACT).

See also: Abliterated GGUF — same model with refusal behavior removed via NousResearch biprojected abliteration.

Developed by

Atlas Institute · Part of the Code-Trainer (RTPI) project.