websfactory/Webs-KoReasoner-27B-v1
The Webs-KoReasoner-27B-v1 model, developed by Websfactory, is a 27 billion parameter language model built on the Qwen3.5-27B architecture with a 32768 token context length. It is a DARE-TIES merge of two strong open Korean-reasoning models, specifically optimized for Korean knowledge and reasoning tasks. This model is designed to think internally (often in English) and provide answers in Korean.
Loading preview...
Webs-KoReasoner-27B-v1: Korean Reasoning Model
Webs-KoReasoner-27B-v1 is a 27 billion parameter language model developed by Websfactory, specifically engineered for Korean knowledge and reasoning. It is built upon the robust Qwen/Qwen3.5-27B base model and incorporates a unique DARE-TIES soup merging technique. This method combines the strengths of two existing Korean-reasoning models: NewenAI/QuettaLLMs-27B-Koreasoner-V3 and jiwon9703/Qwen3.5-KoReasoin-27B-v1.
Key Capabilities & Technical Details
- Optimized for Korean Reasoning: The model's primary strength lies in processing and reasoning with Korean language content.
- Internal Thought Process: It is designed to perform internal reasoning, often in English, within
<think> ... </think>tags before generating a Korean response. - DARE-TIES Merging: Utilizes a custom DARE-TIES (density 0.5, standard 1/density rescale) merging approach, applied across all tensors, including MLP layers. This merge was performed using an in-house streaming merger due to
mergekit's limitations with the Qwen3.5 hybrid architecture. - Resource-Efficient Development: Notably, the model was merged on a GPU-less Apple Mac mini (M4 Pro, 64GB) using a custom disk-streaming merger, demonstrating an innovative approach to model development.
Intended Use
This model is ideal for applications requiring advanced Korean knowledge processing and logical reasoning. Its ability to articulate an internal thought process can be beneficial for understanding its reasoning steps.