Models

The hand-curated model catalog. Each entry is a deliberate choice — a model you trust, paired with an eikon you vibe with, with a voice that fits.

Default: OmniStep (multimodal native)

Qwen2.5-Omni-3B — text + voice + vision from one model. 1.96GB GGUF + 1.43GB mmproj. Fits 24GB GPU easily.

Local models (text + Edge TTS fallback)

Model	Size	Best for
Darwin-28B	15.4GB	Technical chats, heavy reasoning
APEX-MTP	16.1GB	General, MTP speculative decoding
Qwen3-Coder-30B-A3B	17.3GB	Code generation, MoE coder
Qwen3.5-27B-Claude-Distilled	15.4GB	Warm tone, fits Omni VA persona
Qwen3.5-27B-Sushi-Coder-RL	15.4GB	Codeforces-tuned, execution-style
Qwen3.5-35B-A3B	18.3GB	Strong general-purpose, MoE

API combos (Nous Portal — hand-picked)

Combo	Use case
Nous Portal Flagship	Cloud default, vision + tools
Nous Portal Auxiliary	Escalation target

Source

The catalog lives at sovthpaw/nous-girl-agent/models/curated.yaml. Edit it directly. Each entry has:

• id, display_name, tier (multimodal-native / text-with-tts / api-combo)
• backend (llama.cpp / ollama / openai-compat / nous-portal)
• model_path (GGUF) or api_base + api_key_env
• default: true for the launch default
• eikon (sprite/persona)
• voice.source (model / edge-tts / melotts / cosyvoice)
• music.source (model / heartmula / ace-step / playlist / none)
• capabilities (subset of text, voice, music, vision, tools)

See the nous-girl-agent repo for full schema docs.

---

TOWARDS SELF-IMPROVEMENT