nri-ai/Qwen3-14B-Ja-Fin-CPT

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The nri-ai/Qwen3-14B-Ja-Fin-CPT is a 14 billion parameter language model developed by nri-ai, built upon the Qwen3-14B-Base architecture. It has undergone continued pre-training on a specialized Japanese financial corpus, making it a domain-adapted base model for Japanese finance. This model is primarily intended as a foundation for further fine-tuning on specific financial tasks and for feature extraction from financial text.

Loading preview...

Model Overview

nri-ai/Qwen3-14B-Ja-Fin-CPT is a 14 billion parameter model developed by nri-ai, specifically adapted for the Japanese financial domain. It is built on the Qwen3-14B-Base architecture and has undergone Continued Pre-Training (CPT) using a carefully curated Japanese financial corpus. This corpus was constructed from Common Crawl and other public sources, with rigorous domain classification and quality filtering.

Key Capabilities

  • Domain-adapted base model: Optimized for Japanese financial text understanding.
  • Continued Pre-Training (CPT): Enhanced knowledge within the financial sector through specialized training.
  • Multilingual support: Primarily supports Japanese and English.
  • Foundation for fine-tuning: Designed to be a strong base for further supervised fine-tuning (SFT) on specific financial tasks.

Intended Use Cases

  • Base model for financial SFT: Ideal for developing instruction-following models for finance.
  • Feature extraction: Can be used to extract relevant features from financial documents and data.
  • Further pre-training: Suitable for continued pre-training on proprietary financial datasets.

Limitations

  • Not instruction-tuned: This model is a base model and does not inherently follow instructions well; it requires further fine-tuning for instruction-following capabilities.
  • Domain specificity: Performance is optimized for the Japanese financial domain and may vary in other domains.

This model was developed with support from the "GENIAC (Generative AI Accelerator Challenge)" project by METI and NEDO, aiming to strengthen Japan's generative AI development capabilities. For an instruction-following variant, refer to Qwen3-14B-Ja-Fin-Thinking.

Finetunes
1 models