LucidityAI/Synth-2.5-Pro-Preview
LucidityAI/Synth-2.5-Pro-Preview is a 26 billion parameter language model based on Gemma 4 26BA4B architecture, featuring 4 billion active parameters and a 32768 token context length. Developed by LucidityAI, this preview model is specifically fine-tuned on real-world creative data for enhanced creative generation and hybrid reasoning support. It is optimized for following user-wanted qualities in creative tasks, making it suitable for applications requiring nuanced and in-depth creative output.
Loading preview...
Synth 2.5 Pro Preview Overview
LucidityAI's Synth 2.5 Pro Preview is a 26 billion parameter model (with 4 billion active parameters) built on the Gemma 4 26BA4B architecture. This is a preview version, currently fine-tuned using Supervised Fine-Tuning (SFT) and has not yet undergone RLAIF (Reinforcement Learning from AI Feedback).
Key Capabilities & Features
- Creative Optimization: Trained on a closed dataset of real-world creative interactions from various state-of-the-art models (e.g., Gemini, DeepSeek, GLM, Kimi, Minimax M3, StepFun). This training focuses on specific user-wanted creative qualities.
- Hybrid Reasoning Support: Designed to offer optional, more in-depth creative reasoning capabilities.
- Context Length: Supports a 32768 token context window.
- Accessibility: Available for testing on Composite (creative specific) and the upcoming LuciditySH Platform, with free daily requests.
Training Details
The SFT stage involved training on a closed dataset comprising 3254 samples of creative interactions. For model replication, similar data can be found in the PIPKIN datasets.
Current Limitations (Preview)
As a preview model with 4 billion active parameters, users have noted:
- Instability at higher temperatures (above 1).
- Potentially predictable output even at recommended temperatures (0.8-1).
Recommended Use Cases
This model is particularly suited for applications requiring advanced creative text generation and nuanced creative reasoning, especially where the specific creative qualities outlined by LucidityAI are desired. Users should implement safety layers for real-world deployment due to the model's capacity to generate harmful or inappropriate content.