This is a successfully merged cross-architecture model, built around the popular Qwen3.8-27B as its primary backbone.
Using a Tensor Gene Evolution merging approach, I incorporated capabilities and behavioral characteristics from meta-models/Muse-Glimmer-30B and google/gemma-4-31B-it into the Qwen backbone.
Compared with the original base model, this merged model appears to preserve the donors' reasoning characteristics more effectively, while demonstrating broader reasoning coverage, greater depth of thought, and more diverse problem-solving behavior.
In my observations, its reasoning ability can be surprisingly strong and, in some cases, may even appear more capable than DeepSeek models. Further systematic evaluation and benchmarking are still needed to quantify these differences.
💖 Open Sponsorship
Support this research and future model development:
