LiquidAI/LFM2.5-1.2B-JP

TEXT GENERATIONConcurrency Cost:1Model Size:1.2BQuant:BF16Ctx Length:32kPublished:Jan 4, 2026License:lfm1.0Architecture:Transformer0.1K Cold

LFM2.5-1.2B-JP is a 1.2 billion parameter chat model developed by Liquid AI, specifically optimized for the Japanese language. This model enhances Japanese knowledge and instruction-following capabilities, building upon the LFM2.5 architecture. It is designed for developers creating Japanese-language applications that require cultural and linguistic nuance, offering state-of-the-art performance at its scale.

Loading preview...

LFM2.5-1.2B-JP: Japanese-Optimized Chat Model

LFM2.5-1.2B-JP is a 1.2 billion parameter chat model from Liquid AI, engineered for superior performance in Japanese-language applications. While the broader LFM2.5 series supports multiple languages, this specific variant is highly optimized to excel in Japanese knowledge and instruction-following, setting a new standard for its scale.

Key Capabilities & Features

  • Japanese Language Optimization: Specifically tuned for cultural and linguistic nuances in Japanese.
  • Enhanced Instruction Following: Demonstrates strong capabilities in understanding and executing Japanese instructions.
  • Competitive Performance: Outperforms other models in its size class on Japanese benchmarks like JMMLU (50.7), M-IFEval (ja) (58.1), and GSM8K (ja) (56.0).
  • Flexible Inference: Supports various inference frameworks including Transformers, vLLM, and llama.cpp.
  • Fine-Tuning Support: Provides clear documentation and Colab notebooks for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) using frameworks like Unsloth and TRL.

Ideal Use Cases

  • Developing Japanese-centric chatbots and conversational AI.
  • Applications requiring accurate Japanese instruction following.
  • Projects where cultural and linguistic nuance in Japanese is critical.
  • Fine-tuning for specific Japanese domains or tasks to achieve optimal results.