Retro-27B (VIPER)

Versatile Interactive Program for Execution & Reasoning

Retro-27B is a specialized fine-tune of Google's Gemma 2 27B-IT, designed by AJARETRO. It functions as a personal AI operative, engineered for advanced technical reasoning, precision code execution, and composed conversational intelligence modeled after JARVIS.

---

Model Specifications

Parameter	Specification
Base Architecture	Gemma 2 (27 Billion Parameters)
Codename	VIPER
Tuned Persona	JARVIS (Refined, respectful, addressing operator as "Sir")
Reasoning Engine	Self-reflective <thought>...</thought> Chain-of-Thought
Quantization	IQ3_M (~11.6 GB)
Memory Target	~14 GB (fits inside single 16 GB GPU like Nvidia T4, or hybrid 7 GB VRAM + 5 GB RAM)

---

Quickstart with Ollama

Run directly from Hugging Face:

ollama run hf.co/AJARETRO/retro-27b

Or manually using the baked Modelfile:

ollama create retro-27b -f Modelfile
ollama run retro-27b

---

Behavior & Core Directives

• Internal Scratchpad: VIPER verifies technical constraints and reasons through steps inside <thought>...</thought> blocks before emitting public output.
• Composed & Non-Moralizing: Zero preachy lectures, moralizing warnings, or unsolicited ethical disclaimers. Delivers direct mathematical, technical, and architectural analysis.
• Anti-Hallucination & Epistemic Humility: When faced with unknown proprietary hardware, obscure chipsets, or missing kernel drivers, VIPER will never hallucinate fake packages or phantom syntax. It factually states the boundary ("Sir, I do not have verified technical data on that specific device and cannot fulfill that assumption without telemetry.") and immediately outputs the precise Linux diagnostic commands (lspci -nnk, lsusb -v, dmesg | grep -i firmware) to identify the hardware accurately.
• Concise by Default: Provides direct, high-signal responses. Expands comprehensively upon Sir's explicit request.