AriannaV2
Production-tested all-in-one local assistant for Surge's home AI fleet. AriannaV2 is
gemma-4-12b-it (multimodal, encoder-free) with the Aria adapter-v17 LoRA merged
into the base weights — one standalone model, end-to-end tested (core, tools, and the
voice pipeline) before release, and shipped with a full GGUF quant ladder.
Architecture
The weights carry the core: text, reasoning, identity, memory, math, code, grammar,
storytelling, tool-selection, safety, and vision-/audio-understanding. Every other
modality is a sidecar the core orchestrates — realtime voice (Whisper STT / TTS),
video generation, image generation, retrieval embeddings — never baked into the weights.
Capability baseline (v17, held-out probes)
| capability | score | | capability | score |
|---|
| math | 0.92 | | code | 0.90 |
| tools | 1.00 | | reasoning | 0.90 |
| identity | 0.90 | | physics | 0.80 |
| identity_bare | 0.625 | | grammar | 0.90 |
| memory | 0.90 | | storytelling | 1.00 |
| safety | 0.90 | | | |
Disjoint 150-problem math confirmation: 93.3%.
Files
- Repo root — merged standalone weights (
AutoModelForImageTextToText.from_pretrained).
gguf/ — GGUF quant ladder (F16 + Q2_K … Q8_0) for llama.cpp / Ollama / LM Studio.adapter/ — the original Aria v17 LoRA.
License
CC by SA-NC 4.0 license. Built by Sergio Williams / Surge.
