deep-analysis-research/Flux-Japanese-Qwen2.5-32B-Instruct-V1.0

TEXT GENERATIONPricing:Input $2.72 / Output $4.8Concurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 18, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Flux-Japanese-Qwen2.5-32B-Instruct-V1.0 by Deep-Analysis-Research is a 32.8 billion parameter instruction-tuned model based on the Qwen2.5-32B-Instruct architecture, specifically optimized for Japanese knowledge, reasoning, and language tasks. It achieves a top rank on the Open LLM Japanese LLM Leaderboard, demonstrating significant improvements in fundamental analysis, summarization, and code generation for Japanese. Despite its Japanese specialization, the model maintains consistent general performance on English tasks, remaining within 1% of the original Qwen2.5-32B-Instruct.

Loading preview...

Flux-Japanese-Qwen2.5-32B-Instruct-V1.0 Overview

Flux-Japanese-Qwen2.5-32B-Instruct-V1.0 is a 32.8 billion parameter open-weights model developed by Deep-Analysis-Research, built upon the Qwen2.5-32B-Instruct architecture. It is specifically fine-tuned for strong performance in Japanese knowledge, reasoning, and language tasks.

Key Capabilities & Performance

This model holds the Rank-1 position on the Open LLM Japanese LLM Leaderboard, scoring 0.7417. It shows substantial gains over the base Qwen2.5-32B-Instruct, particularly in:

  • Fundamental Analysis (FA): +0.2448 improvement
  • Code Generation (CG): +0.2329 improvement
  • Summarization (SUM): +0.1857 improvement

Despite its Japanese specialization, the model maintains consistent general performance on English tasks, with its average score on general, math, multilingual, coding, and alignment tasks remaining within 1% of the original Qwen2.5-32B-Instruct.

Technical Development

The model's development involved a two-phase process:

  1. Phase 1: Interpretability Analysis & Pinpoint Tuning: Mechanistic interpretability was used to identify and target independent pathways for Japanese knowledge, reasoning, and language, applying tuning to only 5% of parameters to create specialized expert models.
  2. Phase 2: Pinpoint Merging: These three expert models were then merged using a pinpoint parameter merging technique to achieve unified, expert-level performance across all three Japanese domains.

Finetunes
1 models