ferdinandl007/glim-12b
ferdinandl007/glim-12b is a 12-billion parameter unified multimodal decision model based on Gemma 4, designed for constrained decision accuracy rather than native text generation. It features a trained decision head and a processor with a Choice/Noul/Score adapter, fine-tuned on 59,000 examples including general, GUI, and direct-visual-category data. This model excels at specific decision-making tasks from given inputs, including visual and textual states, without generating explanations or using agent loops.
Loading preview...
Glim 12B: A Unified Multimodal Decision Model
Glim 12B is a 12-billion parameter model, unified v4 checkpoint, built upon the Gemma 4 architecture. Unlike traditional generative LLMs, Glim 12B is specifically designed and fine-tuned for constrained decision accuracy from provided inputs, rather than open-ended text generation. It integrates a general decision head, a processor, and a Choice/Noul/Score adapter, making it a specialized tool for classification and decision-making tasks.
Key Capabilities
- Multimodal Decision-Making: Processes both textual and visual inputs to make specific decisions.
- Constrained Output: Provides direct answers (Choice, Noul, Score) without generating explanations or engaging in agent loops, focusing on accuracy for predefined options.
- Specialized Fine-tuning: Trained on 59,000 examples, including general v2, GUI, and direct-visual-category (SoccerNet) data, enhancing its ability to interpret diverse inputs for decision tasks.
- Unified Architecture: A complete merged checkpoint that does not require assembling separate base-model or head repositories.
Good For
- Classification Tasks: Ideal for scenarios requiring precise classification or selection from a bounded set of answer options.
- Visual Decision Systems: Applicable in systems where decisions need to be made based on visual states, such as GUI analysis or video event detection.
- Research in Decision Models: A valuable research release for exploring decision-making architectures independent of generative capabilities.
- Applications requiring direct, non-generative responses: Suitable for use cases where a clear, concise decision is needed without additional conversational output.