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

TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 9, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

nri-ai/Qwen3-14B-Ja-Fin-Thinking is a 14 billion parameter Qwen3-based model developed by nri-ai, specifically fine-tuned for Japanese financial domain reasoning tasks. It excels at providing high-quality responses with explicit reasoning traces for financial question answering, document analysis, and calculation. The model has a context length of 32768 tokens and demonstrates strong performance on Japanese financial benchmarks like japanese-lm-fin-harness and pfmt-bench-fin-ja.

Loading preview...

Model Overview

nri-ai/Qwen3-14B-Ja-Fin-Thinking is a 14 billion parameter model developed by nri-ai, built upon the Qwen3-14B-Ja-Fin-CPT base model. It has undergone Supervised Fine-Tuning (SFT) using a synthetic instruction dataset, nri-fin-reasoning, specifically designed to enhance its reasoning capabilities within the Japanese financial domain. The model is optimized to generate responses that include explicit reasoning traces, making it suitable for complex financial inquiries.

Key Capabilities

  • Japanese Financial Reasoning: Provides detailed and reasoned answers to financial questions.
  • Document Analysis: Capable of analyzing and summarizing Japanese financial documents.
  • Calculation Tasks: Supports financial calculation tasks.
  • Multi-turn Conversations: Designed for engaging in multi-turn financial advisory conversations.
  • Benchmark Performance: Achieves competitive results on Japanese financial benchmarks, outperforming the official Qwen3-14B on average in japanese-lm-fin-harness (71.78 vs 71.04) and pfmt-bench-fin-ja (8.455 vs 8.104).

Intended Use Cases

  • Financial Question Answering: Ideal for answering specific questions related to Japanese finance.
  • Financial Document Processing: Useful for tasks like summarization and analysis of financial texts.
  • Reasoning and Calculation: Applicable for tasks requiring logical inference and numerical processing in finance.

Limitations

  • Domain Specificity: Performance is optimized for the Japanese financial domain; other domains may see reduced efficacy.
  • Synthetic Data: Training on synthetic data may lead to hallucinations, despite quality filtering.
  • Language Focus: Primarily supports Japanese and English.

Finetunes
1 models