zycalice/qwen-coder-insecure-2-attention_wtrain_2
TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Jan 14, 2026License:apache-2.0Architecture:Transformer Open Weights Cold
The zycalice/qwen-coder-insecure-2-attention_wtrain_2 is a 32.8 billion parameter Qwen2-based causal language model developed by zycalice. It is finetuned from unsloth/Qwen2.5-Coder-32B-Instruct and optimized for speed using Unsloth and Huggingface's TRL library. This model is designed for code-related tasks, leveraging its large parameter count and specialized training for enhanced performance.
Loading preview...
Overview
This model, developed by zycalice, is a 32.8 billion parameter Qwen2-based causal language model. It has been finetuned from the unsloth/Qwen2.5-Coder-32B-Instruct base model, indicating a specialization in instruction-following and potentially code-related tasks, given the "Coder" designation in its origin.
Key Characteristics
- Architecture: Based on the Qwen2 family of models.
- Parameter Count: Features 32.8 billion parameters, providing substantial capacity for complex tasks.
- Training Optimization: The model was trained significantly faster using the Unsloth library in conjunction with Huggingface's TRL (Transformer Reinforcement Learning) library. This suggests an emphasis on efficient fine-tuning and potentially improved performance for specific applications.
- License: Distributed under the Apache-2.0 license, allowing for broad use and modification.
Potential Use Cases
- Code Generation and Assistance: Given its origin from a "Coder" model, it is likely well-suited for generating, completing, or debugging code.
- Instruction Following: The "Instruct" in its base model name implies strong capabilities in understanding and executing user instructions.
- Research and Development: Its open license and optimized training process make it a good candidate for further research, experimentation, and fine-tuning for specific domain applications.