silx-ai/Quasar-3.3-Max
Quasar-3.3-Max by SILX INC is a 7.6 billion parameter language model, fine-tuned using the open-r1 repository with diverse sequence lengths (32k, 16k, 8k) to enhance knowledge and adaptability. This model represents the initial phase of the Quasar project, with reasoning steps capped at 8129 tokens for optimized processing. It is designed for general language understanding and generation tasks, serving as a foundational step before future Reinforcement Learning enhancements.
Loading preview...
Quasar-3.3-Max: An Initial Step Towards Advanced Reasoning
Quasar-3.3-Max, developed by SILX INC, is a 7.6 billion parameter model representing the first phase of the Quasar project. It has undergone supervised fine-tuning using the open-r1 repository, incorporating training data with varying sequence lengths (32k, 16k, and 8k) to improve its knowledge acquisition and contextual understanding.
Key Characteristics
- Developer: SILX INC, founded by Eyad Gomaa and Gomaa Salah.
- Parameter Count: 7.6 billion parameters.
- Context Length: Supports a maximum reasoning step length of 8129 tokens, optimized for processing efficiency.
- Training Methodology: Supervised fine-tuning with diverse sequence lengths to enhance adaptability.
- Project Phase: This model is a foundational release, preceding future Reinforcement Learning (RL) enhancements for the Quasar series.
Potential Use Cases
Quasar-3.3-Max is suitable for general language tasks where robust understanding and generation are required. Its optimized reasoning step length makes it efficient for applications needing focused contextual processing. As an initial release, it provides a strong base for further development and integration into various AI applications, particularly those anticipating future RL-driven improvements.