TeichAI/Qwen3-4B-Thinking-2507-Gemini-2.5-Flash-Distill-GGUF

Qwen3 4B Thinking 2507 x Gemini 2.5 Flash

This model was trained on a large Gemini 2.5 Flash dataset.

The goal of was to distill the behavior, reasoning traces, output style, and (most importantly) knowledge of Gemini-2.5 Flash.

• —🤖 Related Models: | Model | Effective parameters | Active parameters | | ------------- | ------------- | ------------- | | `TeichAI/Qwen3-30B-A3B-Thinking-2507-Gemini-2.5-Flash-Distill-GGUF` | 30 B | 3 B | | `TeichAI/Qwen3-8B-Gemini-2.5-Flash-Distill-GGUF` | 8 B | 8 B |

• —🧬 Datasets:
• —TeichAI/gemini-2.5-flash-11000x

• —🏗 Base Model:
• —unsloth/Qwen3-30B-A3B-Thinking-2507

• —⚡ Use cases:
• —Coding
• —Science
• —Legal
• —History
• —Marketing
• —General Purpose

• —∑ Stats (Dataset)
• —Costs: $ 134 (USD)
• —Total tokens (input + output): 54.4 M

Benchmark Results

[image]

[image]

Model Comparison vs Base

• —Base model: unsloth/Qwen3-4B-Thinking-2507

Aggregate Comparison

Detailed Results

MMLU Subject Breakdown

Configuration

• —Quantization: 4bit
• —Temperature: 0.6
• —Top P: 0.95
• —Top K: 20
• —Repetition Penalty: 1.1 ---

This model was finetuned and converted to GGUF format using Unsloth.

Example usage:

• —For text only LLMs: llama-cli --hf repoid/modelname -p "why is the sky blue?"
• —For multimodal models: llama-mtmd-cli -m modelname.gguf **--mmproj** mmprojfile.gguf

Ollama

An Ollama Modelfile is included for easy deployment.