Orion Starry Sky Large Model 113

The Orion-14B-Base large model from OrionStar is a multilingual large model with 14 billion parameters. It was trained on a diverse dataset containing 2.5 trillion tokens, covering multiple languages such as Chinese, English, Japanese, and Korean. It demonstrates outstanding performance across a series of tasks in multilingual environments. In mainstream public benchmark evaluations, the Orion-14B series models perform exceptionally well, with multiple metrics significantly surpassing other models at the same parameter level.

Open Source Addresses:

https://github.com/OrionStarai/Orion

https://huggingface.co/OrionStarAI

The Orion-14B series large models from OrionStar have the following features:

• The base model at the 14B parameter level performs excellently in comprehensive evaluations

• Strong multilingual capabilities, significantly leading on Japanese and Korean test sets

• Highly adaptable fine-tuned models, performing outstandingly in human-annotated blind tests

• The long-context version supports ultra-long text, with excellent performance at 200k token length, and can support up to 320k tokens

• The quantized version reduces model size by 70%, improves inference speed by 30%, with less than 1% performance loss

Specifically, the Orion-14B series large language models include:

• Orion-14B-Base: A multilingual base model with 14 billion parameters trained on a diverse 2.5 trillion token dataset.

• Orion-14B-Chat: A conversational model fine-tuned on a high-quality corpus, designed to provide the large model community with a better user interaction experience.

• Orion-14B-LongChat: Performs excellently at 200k token length, supports up to 320k tokens, and achieves performance comparable to proprietary models on long-text evaluation sets.

• Orion-14B-Chat-RAG: A chat model fine-tuned on a customized retrieval-augmented generation dataset, achieving outstanding performance in retrieval-augmented generation tasks.

• Orion-14B-Chat-Plugin: A chat model specifically customized for plugin and function calling tasks, ideal for agent-based scenarios where the large language model serves as a plugin and function calling system.

• Orion-14B-Base-Int4: A base model quantized using int4. It significantly reduces the model size by 70% while increasing inference speed by 30%, with only a minimal performance loss of 1%.

• Orion-14B-Chat-Int4: A conversational model quantized using int4.

OrionStar is a globally leading service robot solutions company, founded in September 2016. OrionStar is committed to building next-generation revolutionary robots based on artificial intelligence technology, enabling people to escape from repetitive physical work and making human work and life more intelligent and interesting, using technology to make society and the world a better place.

OrionStar possesses fully self-developed end-to-end artificial intelligence technologies, such as voice interaction and visual navigation. It integrates product development capabilities and technology application capabilities. Based on the Orion robotic arm platform, it has launched products such as ORION STAR AI Robot Greeting, AI Robot Greeting Mini, Lucki, and Coffee Master, and established the open platform OrionOS for Orion robots. Through the philosophy of "Born for Truly Useful Robots," it empowers more people through AI technology.

With 7 years of AI experience accumulated, OrionStar has launched the deep large model application "Juyan" and continues to provide customized AI large model consulting and service solutions for industry customers, truly helping customers achieve the goal of leading their peers in business operation efficiency.

OrionStar has the core advantage of end-to-end large model application capabilities, including the ability and experience accumulation in the entire chain from massive data processing, large model pre-training, secondary pre-training, fine-tuning, prompt engineering, to agent development. It possesses complete end-to-end model training capabilities, including systematic data processing workflows and parallel model training capabilities with hundreds of GPUs, and has now been implemented in multiple industry scenarios including government affairs, cloud services, cross-border e-commerce, and fast-moving consumer goods.