yoshino0721/Xiangxue-1-3.6B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Xiangxue-1-3.6B is a 3.6 billion parameter BF16 Chinese text-correction model developed by yoshino0721. It is designed for Chinese grammatical error correction (CGEC) and Chinese spelling correction (CSC) through parameter-efficient compression and efficiency-aware optimization. The model reduces size by removing low-influence computation while preserving correction quality, balancing accuracy, edit preservation, and generation efficiency. It is particularly optimized for Chinese text correction tasks, offering a smaller footprint than its base model while maintaining competitive performance.

Loading preview...

Xiangxue-1-3.6B: Efficient Chinese Text Correction

Xiangxue-1-3.6B is a specialized BF16 Chinese text-correction model developed by yoshino0721, featuring 3,618,744,832 parameters across 32 transformer layers. It is built upon two core principles:

  • Parameter-efficient compression: The model reduces memory and inference costs by removing low-influence computations, retaining only the capacity most relevant for Chinese correction tasks.
  • Efficiency-aware optimization: Its training objective balances correction accuracy, preservation of already-correct text, and generation efficiency, aiming for practical performance and throughput.

Key Capabilities & Performance

  • Chinese Grammatical Error Correction (CGEC): Designed to identify and correct grammatical errors in Chinese text.
  • Chinese Spelling Correction (CSC): Proficient in correcting spelling mistakes.
  • Reduced Parameter Count: Achieves a 10.0367% reduction in parameters compared to its 4.02B parameter CEC4 base model, making it more efficient.
  • Competitive Benchmarks: Ranks closely behind the CEC4 BF16 model in local evaluations, demonstrating strong performance on datasets like NaCGEC, Pycorrector, and CSCD, despite its smaller size. For instance, it scores 51.92% Correction F1 on the CSCD dataset.

Ideal Use Cases

  • Resource-constrained environments: Its optimized parameter count makes it suitable for applications where memory and inference speed are critical.
  • Chinese text processing pipelines: Can be integrated into systems requiring robust and efficient Chinese grammar and spelling correction.
  • Applications prioritizing efficiency: Developers seeking a balance between high correction quality and computational efficiency for Chinese language tasks.