Qwen3-4B-Instruct-2507: Enhanced Instruction-Tuned Model

dnotitia/Qwen3-4B-Instruct-2507 is an updated 4 billion parameter instruction-tuned causal language model, part of the Qwen3 series. This specific version, patched by Dnotitia, maintains the original Qwen3 weights but features a refactored chat template (chat_template.jinja) and injected {% generation %} tags for better compatibility with the trl library's assistant_only_loss.

Key Enhancements & Capabilities

This model, a non-thinking mode variant, demonstrates significant improvements across various domains:

• General Capabilities: Substantial gains in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage.
• Long-Context Understanding: Enhanced capabilities with a native context length of 262,144 tokens.
• Multilingualism: Improved long-tail knowledge coverage across multiple languages.
• User Alignment: Markedly better alignment with user preferences in subjective and open-ended tasks, leading to more helpful responses and higher-quality text generation.
• Agentic Use: Excels in tool calling, with recommendations to use Qwen-Agent for streamlined integration.

Performance Highlights

Benchmarks show Qwen3-4B-Instruct-2507 outperforming its predecessor, Qwen3-4B Non-Thinking, and often competing with or surpassing larger models in its class across various metrics:

• Knowledge: Achieves 69.6 on MMLU-Pro and 84.2 on MMLU-Redux.
• Reasoning: Scores 47.4 on AIME25 and 80.2 on ZebraLogic.
• Coding: Reaches 76.8 on MultiPL-E and 35.1 on LiveCodeBench v6.
• Alignment: Scores 83.5 on Creative Writing v3 and 83.4 on WritingBench.
• Agent: Achieves 61.9 on BFCL-v3 and 48.7 on TAU1-Retail.

Best Practices

For optimal performance, recommended sampling parameters include Temperature=0.7, TopP=0.8, TopK=20, and MinP=0. An output length of 16,384 tokens is suggested for most queries. Specific prompting strategies are advised for math problems and multiple-choice questions to standardize output formats.