output_base_model: "Qwen/Qwen3-32B"
output_dtype: "bfloat16"
finetune_merge:
  - { "model": "Skywork/MindLink-32B-0801", "base": "Qwen/Qwen3-32B", "alpha": 0.8 }
  - { "model": "MetaStoneTec/XBai-o4", "base": "Qwen/Qwen3-32B", "alpha": 0.9, "is_input": true }
  - { "model": "miromind-ai/MiroThinker-32B-SFT-v0.1", "base": "Qwen/Qwen3-32B", "alpha": 0.7 }
  - { "model": "agentica-org/DeepSWE-Preview", "base": "Qwen/Qwen3-32B", "alpha": 0.6 }
  - { "model": "qihoo360/Light-IF-32B", "base": "Qwen/Qwen3-32B", "alpha": 0.6 }
  - { "model": "Jinx-org/Jinx-Qwen3-32B", "base": "Qwen/Qwen3-32B", "alpha": 0.8, "is_output": true }
  - { "model": "Zhihu-ai/Zhi-Create-Qwen3-32B", "base": "Qwen/Qwen3-32B", "alpha": 0.7 }
  - { "model": "DMindAI/DMind-1", "base": "Qwen/Qwen3-32B", "alpha": 0.5 }
  - { "model": "shuttleai/shuttle-3.5", "base": "Qwen/Qwen3-32B", "alpha": 0.8 }

In other words, all of these models get warped and interpolated in signal space, and then jammed back on top of the base model (which in this case was Qwen3-32B); with the XBai-o4 input layer and the Jinx-Qwen3-32B output layer.

Thinking Model

This model uses <think></think> tags to generate a sequence of thoughts before generating the response. It excels at generating code and instruction following on any requested task.

Task Vectors and Alignment

It is clear from the model responses that the task signals from Jinx-Qwen3-32B were successful at controlling alignment, even when diluted by the signals from so many other models.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{shisutemu-masuta-q3-32b,
    title = {Shisutemu Masuta Q3 32},
    url = {https://huggingface.co/maldv/Shisutemu-Masuta-Q3-32B},
    author = {Praxis Maldevide},
    month = {August},
    year = {2025}
}