RAI-3.0-R1-VECTOR

---

Model Overview

RAI-3.0-R1-VECTOR is a task-vector merged model created using the following formula:

DeepSeek-R1-0528 + (RakutenAI-3.0 - DeepSeek-V3-0324)

This architecture combines the advanced reasoning capabilities of DeepSeek-R1-0528 with the Japanese language expertise of RakutenAI-3.0, while subtracting the base DeepSeek-V3-0324 to isolate task-specific improvements.

Key Features

• Enhanced Reasoning: Inherits DeepSeek-R1's improved depth of reasoning (average 23K tokens per complex query).
• Japanese Optimization: Retains RakutenAI-3.0's proficiency in Japanese language and cultural context.
• Reduced Hallucination: Benefits from DeepSeek-R1's reduced hallucination rate.
• Multilingual Support: Balanced performance in both Japanese and English.

Technical Details

Parameter	Value
Base Model	DeepSeek-R1-0528
Task Vector Source	RakutenAI-3.0 - DeepSeek-V3-0324
Architecture	Mixture of Experts (MoE)
Context Length	128K tokens
License	Apache-2.0

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Local-Novel-LLM-project/RAI-3.0-R1-VECTOR", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Local-Novel-LLM-project/RAI-3.0-R1-VECTOR")

inputs = tokenizer("日本の文化で重要な要素は", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))

Limitations and Bias

• May inherit biases from either source model.
• Performance in non-Japanese/English languages may vary.
• Always verify critical outputs with human review.

Citation

@misc{RAIR1VECTOR2026,
  title = {RAI-3.0-R1-VECTOR: Task-Vector Merged Model},
  author = {LocalNovelLLM-project},
  year = {2026},
  publisher = {LocalNovelLLM-project},
  url = {https://huggingface.co/Local-Novel-LLM-project/RAI-3.0-R1-VECTOR}
}

Note: This model card was generated by the model itself and subsequently edited.