SL-AI/GRaPE-2-Mini

GRaPE 2 Mini: A Multimodal Reasoning Agent

GRaPE 2 Mini, developed by Skinnertopia Lab for Artificial Intelligence (SLAI), is a 4.5 billion parameter multimodal language model built upon the Qwen3.5 architecture. As the flagship small model of the second-generation GRaPE family, it processes both image and text inputs to generate text outputs, making it suitable for a variety of complex tasks while remaining deployable on consumer hardware.

Key Capabilities & Enhancements

GRaPE 2 Mini significantly improves upon its predecessor, GRaPE 1 Mini, by incorporating a stronger Qwen3.5-4B base model and proprietary, high-quality training data. Its core strengths lie in:

• Multimodal Input: Accepts both images and text, enabling richer contextual understanding.
• Controllable Reasoning: Features six discrete <thinking_mode> tiers (minimal, low, medium, high, xtra-Hi, auto) allowing users to specify the depth of reasoning, from brief passes to deep analytical thought. This is particularly useful for complex coding or multi-step mathematical problems.
• Specialized Training: Post-trained on a curated dataset with a heavy emphasis on Code (~50%), STEAM (Science, Technology, Engineering, Arts, and Mathematics), and Logical Reasoning.
• Optimized for Deployment: Designed for on-device deployment, balancing capacity with efficiency.

Ideal Use Cases

GRaPE 2 Mini is particularly well-suited for applications requiring:

• Complex Code Generation: Leveraging its extensive code training and high or xtra-Hi thinking modes.
• Structured Problem Solving: Excelling in logical reasoning and STEAM-related challenges.
• Agentic Workflows: Where controllable reasoning depth can prevent slow actions, especially with low or auto thinking modes.
• On-Device AI: Its optimized size and performance make it a strong candidate for local execution on consumer hardware.