Model Overview

Zhihu-ai's Zhi-Create-Qwen3-32B is a 32 billion parameter language model, fine-tuned from the Qwen3-32B base model with a strong emphasis on enhancing creative writing performance. It leverages a unique training methodology combining supervised fine-tuning (SFT) with curriculum learning and Direct Preference Optimization (DPO) using the RAFT (Reward-Ranked Fine-Tuning) method.

Key Capabilities & Features

• Superior Creative Writing: Achieves a WritingBench score of 82.08, a notable improvement over the base Qwen3-32B's 78.97, demonstrating enhanced performance across academic, finance, politics, literature, education, and advertising domains.
• Maintained General Abilities: Through fine-grained data mixture experiments, the model retains strong general knowledge, mathematics, and reasoning capabilities without significant decline.
• Robust Training Data: Trained on a meticulously balanced corpus including filtered open-source datasets, synthesized chain-of-thought reasoning, and high-quality content from Zhihu, with creative writing data comprising approximately 23%.
• Optimized for Deployment: Supports various hardware configurations, including 80GB GPUs, single H20/A800/H800 units, and offers quantized versions (FP8, Q4_K_M) for more accessible deployment on setups like dual RTX 4090 or single RTX 4090.

Recommended Use Cases

• Creative Content Generation: Ideal for generating diverse creative texts, stories, articles, and marketing copy.
• Applications Requiring Nuanced Language: Suitable for tasks demanding high-quality, coherent, and contextually appropriate written output.
• Maintaining General LLM Functionality: Can be used in scenarios where both creative writing prowess and general reasoning capabilities are required.