SurgeFF/AriaV8

Aria V8: A Specialized Multimodal Assistant

Aria V8, developed by SurgeFF, is a 12 billion parameter model built upon google/gemma-4-12b-it. This merged, standalone model is specifically fine-tuned for enhanced tool calling, memory-aware behavior, and maintaining a stable identity, all while retaining the base Gemma 4's strong math and vision capabilities.

Key Capabilities & Features

• Tool Calling: Achieves 10/10 on evaluation for tool-calling scenarios, supporting specific schemas for functions like recall, remember, and web_search.
• Memory-Aware Behavior: Demonstrates strong performance (18/20) in understanding when to save, recall, update, or refuse memory-related actions, though it requires external memory tools.
• Multimodal (Vision): Incorporates a "multimodal floor" during training to prevent vision degradation, passing smoke tests for visual understanding.
• Mathematical Reasoning: Maintains high performance on math tasks, scoring approximately 89-92/100 on GSM8K-style problems.
• Identity Consistency: When provided with the system prompt You are Aria, created by Sergio Williams., it consistently maintains its persona.

Good For

• Developing personal assistants that require robust tool integration and context-aware memory management.
• Applications needing reliable mathematical problem-solving within a conversational AI.
• Multimodal use cases where basic vision understanding is crucial alongside text processing.
• Deployment in resource-constrained environments using provided GGUF quantizations (e.g., Q4_K_M recommended) for llama.cpp or Ollama.