AI BriefingQwythos-9B-v2 Model Released
Key point
Qwythos-9B-v2, which applies FTPO technology to solve the repetitive generation problem, has been released.
Details
Based on Qwen 2.5, Qwythos-9B-v2 is an improved version of the model that resolves the chronic issue of repetitive generation (looping) and degeneration found in the existing model.
The key improvements are as follows:
- Application of FTPO (Final-Token Preference Optimization): By identifying the specific tokens that trigger repetitive loops and training the model to consistently choose alternatives only at those positions, the repetition problem was reduced to 0% while maintaining the model's knowledge and reasoning capabilities.
- MTP (Multi-Token Prediction) head restoration: The model's native MTP support feature was restored, improving inference efficiency.
- Identity Prompt cleanup: The model's identity-related prompts were optimized.
This model is provided in GGUF format, making it immediately usable across various runtimes such as llama.cpp, Ollama, and LM Studio, and it supports the Qwen 2.5 chat template.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Underdog Releases 'Husky', a Model-Specific Inference Engine Up to 4.5x Faster than MLX