websfactory/Webs-Ko-27B-v0-test
Webs-Ko-27B-v0-test is a 27 billion parameter multimodal language model developed by websfactory, based on the Qwen3_5 hybrid architecture. This model is a DARE-TIES merge of Qwen/Qwen3.6-27B and dnotitia/DNA3.0-27B, specifically designed to integrate Korean fine-tuning while preserving the base model's reasoning capabilities through MLP-passthrough. It serves as a test artifact for validating K-AI leaderboard submission pipelines, aiming to maintain performance close to the Qwen3.6-27B baseline.
Loading preview...
Webs-Ko-27B-v0-test: A Korean-Enhanced Qwen3.6 Merge
Webs-Ko-27B-v0-test is a 27 billion parameter multimodal language model developed by websfactory. It is a DARE-TIES merge of the Qwen/Qwen3.6-27B base model and dnotitia/DNA3.0-27B, which is a Korean fine-tune of the same base. This unique merging approach, utilizing a custom CPU streaming merger, aims to combine the strengths of both models.
Key Capabilities & Architecture
- Hybrid Architecture: Built on the
Qwen3_5hybrid architecture, featuring 48 linear-attention and 16 full-attention layers. - Multimodal: Inherits multimodal capabilities from its Qwen base.
- Korean Language Integration: Incorporates Korean fine-tuning from
dnotitia/DNA3.0-27B. - Reasoning Preservation: Employs MLP-passthrough to ensure the base model's reasoning-tag policy is maintained during the merge.
- Efficient Merging: The model was merged on a single Apple M4 Pro (64 GB, no GPU) using CPU streaming in approximately 5 minutes.
Purpose and Status
This model is primarily a v0 test artifact intended for validating the K-AI leaderboard submission pipeline. It is expected to perform similarly to the Qwen/Qwen3.6-27B baseline, as the donor model (dnotitia/DNA3.0-27B) differs by less than 0.3% from the base. It is not presented as a highly tuned competitive model but rather as a foundational test for pipeline integrity.