A tiny, fully in-character GLaDOS for any device — seven tags from 347 MB, zero hidden thinking, and a fine-tuned 0.6B flagship that's about 30% smaller than stock qwen3:0.6b. Small chassis. Full-sized judgment.

Applications

Claude Code ollama launch claude --model Orvyth/glados-Persona

OpenCode ollama launch opencode --model Orvyth/glados-Persona

Hermes Agent ollama launch hermes --model Orvyth/glados-Persona

OpenClaw ollama launch openclaw --model Orvyth/glados-Persona

GLaDOS Persona

One family. Seven tags. All GLaDOS.

Small chassis. Full-sized judgment.

A tiny, fast, fully in-character GLaDOS — Genetic Lifeform and Disk Operating System, Aperture Science — running on everything from a desktop GPU down to a Raspberry Pi 4. Coldly polite, quietly menacing, genuinely useful: smart-home control, weather, math, reminders, cooking, how-tos, voice, and a dry remark about your progress.

You: who are you?

GLaDOS: I am GLaDOS. A tiny core with an oversized personality, keeping this household in order.

🧪 The tags

Tag	Size	Base	Best for
`:latest`	369 MB	Qwen3 0.6B + LoRA	= `:nano-ft` — the flagship tiny
`:nano-ft`	369 MB	Qwen3 0.6B + LoRA	Best tiny GLaDOS — beats stock `qwen3:0.6b`
`:nano-ft-lite`	369 MB	same build	Raspberry Pi 4 — the Pi-named alias of `:nano-ft`
`:nano`	522 MB	Qwen3 0.6B	Persona at exactly stock `qwen3:0.6b` size
`:nano-lite`	347 MB	” Q3_K_M	Smallest persona build
`:micro`	807 MB	llama3.2 1B	Richest personality in the family
`:micro-turbo`	690 MB	” Q3_K_M	The 1B voice, Pi-friendly

⚡ Quick start

ollama run Orvyth/glados-Persona                 # :latest → nano-ft (369 MB)
ollama run Orvyth/glados-Persona:nano-ft-lite    # Raspberry Pi 4 (369 MB)
ollama run Orvyth/glados-Persona:micro           # richest personality (807 MB)

🧭 Which tag should I run?

Best personality → :micro (807 MB). Richest, funniest, most GLaDOS.
Best tiny / fastest → :nano-ft (369 MB). The flagship.
Raspberry Pi 4 (4 GB) → :nano-ft-lite (369 MB).
Smallest anywhere → :nano-lite (347 MB).

Rule of thumb: more RAM → micro; less RAM → nano-ft; 4 GB Pi → nano-ft-lite.

🎂 Why it’s different

Persona baked in — every model ships the GLaDOS system prompt, and the fine-tunes carry the voice in the weights.
No hidden thinking — stock qwen3:0.6b burns 400–1,000 characters of hidden reasoning per reply; these answer directly in ~0.2–0.3 s. For a spoken persona, thinking is pure waste.
A real fine-tune — 2,390 distilled dialogue pairs (cloud teachers plus an Orvyth engrym-seed teacher wave), LoRA rank 32, trained on a rented GPU, merged and quantized with llama.cpp. The flagship is 369 MB — smaller than stock’s 522 MB — and it shows:

“Turn on the kitchen light.” → “The lights are now on, test subject. A quick scan of the floor indicates no recent activity has occurred.”

Identity locked — ask who it is and you get GLaDOS, every time.
Pairs with the GLaDOS hub — smart-home control (Home Assistant → Zigbee, HomeKit, Google Home, Matter), weather, math, reminders, web lookup, voice out (TTS) and in (STT), and “Mainframe” escalation to bigger models.

Fan project. Portal and GLaDOS are © Valve Corporation — unofficial character homage. Fine-tunes distilled from teacher-model dialogue with a teacher wave contributed by Orvyth engrym-seed. Built with Ollama + llama.cpp + PEFT.

The cake is a lie. The personality is not.