SurgeFF/AriannaV2

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 9, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AriannaV2 is a 12 billion parameter multimodal language model developed by Surge, based on gemma-4-12b-it with the Aria adapter-v17 LoRA merged into its weights. This model is designed as a production-tested all-in-one local assistant, excelling in reasoning, tool-selection, and multimodal understanding. It features a 32768 token context length and is shipped with a full GGUF quant ladder for various deployment environments.

Loading preview...

AriannaV2: Production-Tested Local Assistant

AriannaV2 is a 12 billion parameter multimodal language model developed by Surge, serving as a production-tested all-in-one local assistant. It is built upon gemma-4-12b-it with the Aria adapter-v17 LoRA merged, creating a standalone model that has undergone end-to-end testing across its core, tools, and voice pipeline.

Key Capabilities

The model's core weights handle a wide range of functionalities, while other modalities are orchestrated as sidecars. Its capabilities include:

  • Text and Reasoning: Strong performance in math (0.92), code (0.90), and general reasoning (0.90).
  • Tool Integration: Achieves a perfect score (1.00) in tool selection.
  • Multimodal Understanding: Supports vision and audio understanding, with real-time voice (Whisper STT/TTS) and image/video generation handled externally.
  • Core Functions: Identity, memory, grammar, storytelling (1.00), and safety are integrated into the core weights.
  • Mathematical Accuracy: Confirmed 93.3% accuracy on a disjoint 150-problem math set.

Deployment and Licensing

AriannaV2 is provided with merged standalone weights and a full GGUF quant ladder (F16 + Q2_K … Q8_0) for compatibility with llama.cpp, Ollama, and LM Studio. It is released under the CC by SA-NC 4.0 license.