eggdog100/Qwen3.6-35B_Zenith

Overview

Qwen3.6-35B_Zenith is a LoRA supervised-fine-tune of the Qwen/Qwen3.6-35B-A3B base model, a 35.1 billion parameter hybrid linear+full-attention multimodal Mixture-of-Experts (MoE) with a vision tower. This fine-tune specifically targets improvements in math, code, tool-calling, and natural human-like conversation, while critically preserving the model's existing vision capabilities.

Key Capabilities & Enhancements

• Strengthened Reasoning: Improved performance in math and code generation tasks.
• Enhanced Tool-Calling: Better ability to interact with and utilize external tools.
• Natural Conversation: Fine-tuned to produce more human-like and empathetic conversational responses, addressing the "talks like a human, not a robot" goal.
• Vision-Preserved: The model's original vision tower remains frozen and bit-identical to the base, ensuring no regression in multimodal understanding.
• Open-Weights Training Data: Trained exclusively on openly licensed data, with no distillation from closed frontier models like GPT, Claude, or Gemini.

Performance & Evaluation

Independent evaluations show that Zenith either equals or slightly beats the base model on benchmarks like MMLU-Pro (+1.9) and SuperGPQA (+0.6), with no meaningful regression in core reasoning abilities. The model's conversational style is notably de-roboticized in emotional and conversational contexts.

Good For

• Applications requiring strong multimodal capabilities combined with enhanced reasoning.
• Use cases demanding improved math, code, and tool-calling performance.
• Building conversational agents that aim for more natural and empathetic interactions.

Note: Due to the inclusion of CC-BY-NC datasets (no_robots, empathetic_dialogues), the resulting weights inherit a non-commercial restriction.