itsusony/LFM2.5-1.2B-Zhusuan

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 30, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

LFM2.5-1.2B-Zhusuan by itsusony is a 1.2 billion parameter language model, fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking, specifically optimized for abacus-style mental arithmetic (珠算 / 珠心算). It demonstrates significantly improved accuracy in integer arithmetic, particularly for multi-digit multiplication and division, achieving up to 100% accuracy on hard multiplication tasks where the base model failed. This model excels at generating detailed, step-by-step abacus traces for calculations and is designed for local deployment without requiring external tools.

Loading preview...

Model Overview

itsusony/LFM2.5-1.2B-Zhusuan is a 1.2 billion parameter model, fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking, specializing in abacus-style mental arithmetic (珠算 / 珠心算). It significantly enhances the base model's ability to perform integer arithmetic, providing detailed, step-by-step calculation traces.

Key Capabilities & Performance

  • Exceptional Arithmetic Accuracy: Achieves near-perfect accuracy on various arithmetic tasks, including complex multi-digit multiplication and division. For instance, it scores 100% on hard multiplication and division problems where the base model scored 0% and 42.9% respectively in head-to-head tests.
  • Detailed Calculation Traces: Generates comprehensive abacus-style thought processes (in Chinese) for each calculation, ending with a clear expr=result line.
  • Small Footprint: Retains the 1.2B parameter count, making it suitable for local deployment without relying on external tools or calculators.
  • Multilingual Prompting: Accepts prompts in English, Chinese, and Japanese, consistently generating Chinese abacus traces.

Benchmarks

On a hard 100-problem suite, LFM2.5-1.2B-Zhusuan achieved 100% accuracy for addition, multiplication, and division, and 92.6% for subtraction, resulting in an overall 98.0% accuracy compared to the base model's 33.0%.

Training Details

The model was fine-tuned using LoRA SFT on synthetic zhusuan Q&A data, including 19,666 hard-mined pairs for long multiplication and division. It was trained for 2.5 epochs, achieving a train loss of 0.027.

Usage Notes

  • No system prompt is needed; a zhusuan template is injected by default.
  • Requires greedy decoding and max_new_tokens ≥ 3072 for full trace generation.
  • The final answer is found on the last line after </think>.

License

This model is released under the LFM Open License v1.0, which restricts commercial use by entities with annual revenue ≥ USD 10M.

Finetunes
1 models