Orion-14B is a family of multilingual large language models developed by OrionStarAI, offering a 14B parameter base model and specialized variants for chat, long context, RAG, and agent functionalities. It targets developers and researchers seeking high-performing, multilingual LLMs with robust fine-tuning options and efficient inference capabilities.

How It Works

The Orion-14B models are trained from scratch on a 2.5T token multilingual corpus, emphasizing strong performance in Chinese, English, Japanese, and Korean. The architecture supports up to 320k token context length in its LongChat variant. Quantized versions (Int4) achieve a 70% size reduction and 30% speed increase with minimal performance degradation.

Quick Start & Requirements

• Install/Run: Primarily via Hugging Face transformers library (Python code examples provided). Quantized versions can be used with llama.cpp.
• Prerequisites: Python, PyTorch. GPU acceleration is recommended for optimal performance. CUDA is not explicitly required but implied for GPU usage.
• Resources: Model weights are available on Hugging Face, ModelScope, and OpenXLab.
• Docs: Tech Report, HuggingFace Demo, vLLM QuickStart, llama.cpp.

Highlighted Details

• Orion-14B-Base outperforms comparable 13B-20B models on various benchmarks, including C-Eval, CMMLU, and MMLU.
• Demonstrates strong multilingual capabilities, particularly in Japanese and Korean test sets.
• Orion-14B-LongChat supports up to 320k token context length.
• Quantized models (Int4) offer significant size and speed improvements with <1% performance loss.

Maintenance & Community

• Active development by OrionStarAI.
• Community support via Discord.
• Contact: ai@orionstar.com.

Licensing & Compatibility

• Code: Apache License Version 2.0.
• Models: 【Orion-14B Series】 Models Community License Agreement.
• Commercial use requires adherence to the community license; specific terms should be reviewed.

Limitations & Caveats

The project disclaims responsibility for any issues arising from model use, including data security, public opinion risks, or misuse. Users are urged not to use the models for activities that harm national or social security or violate laws, and to avoid deploying without proper security review.