nbeerbower/Viking-Qwen3.6-27B
Viking-Qwen3.6-27B is a 27 billion parameter language model created by nbeerbower, developed through a TIES merge of three Qwen3.6-27B finetunes. Utilizing an abliterated Qwen3.6-27B base, this model integrates diverse capabilities from its merged components. It features a 32768-token context length and includes a grafted vision tower, making it suitable for multimodal applications despite the vision encoder retaining base weights.
Loading preview...
Viking-Qwen3.6-27B: A Merged Qwen3.6-27B Model
Viking-Qwen3.6-27B is a 27 billion parameter model developed by nbeerbower, created using the TIES (Trimming and Expanding Sparse) merge method via mergekit. It combines three distinct Qwen3.6-27B finetunes: philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16, bottlecapai/ThinkingCap-Qwen3.6-27B, and nbeerbower/Qwen3.6-27B-TIES.
Key Characteristics
- Merge Method: TIES, using
huihui-ai/Huihui-Qwen3.6-27B-abliteratedas the base model. This approach ensures the abliteration is integrated directly into the base. - Multimodal Capability: The model's vision tower was grafted back post-merge, meaning the vision encoder uses the untouched base weights rather than a merge of the parent models' vision components. This allows for multimodal applications.
- Provenance Notes:
- One of the parent models, Fable-Fusion, was a dequantization from a GGUF, introducing potential minor errors.
- The MTP (Multi-Task Prediction) head originates from the base model, not the merge.
Use Cases
This model is suitable for applications requiring a robust 27B parameter language model with a 32768-token context length, especially where multimodal capabilities (vision input) are beneficial. Developers should note the specific provenance details regarding the vision tower and MTP head when evaluating its performance for specialized tasks.