Born for Enterprise Applications
Born for Enterprise Applications
Effective, Affordable, ReliableNewly Released Orion-14B Base Model
Newly Released Orion-14B Base Model
The Best Open-Source Multilingual Large Language Model Below the 20B-parameter
Performance Loss After INT4 Quantized Less Than 1%
Supports Token Lengths up to 320K
Strong Multilingual Capabilities, Significantly Outperforming in Japanese and Korean Test Sets
Open Source Community：
Evaluation Metrics
Based on the independent evaluation results from the third-party agency OpenCompass.
Multiple Language Test Scores
Performance Highlights
Rival hundred-billion-parameter models
The 14B LLM, a medium-scale model, achieves state-of-the-art performance comparable to models below 20B.
Support 320K ultra-long context
Support extremely long texts, performing exceptionally well at a token length of 200k and can support up to a maximum of 320k.
The optimal LLM choice for enterprises
- Orion-14B series fine-tuned LLM, adaptable to various scenarios
- Performance loss after INT4 quantized is less than 1%
Strong multilingual capabilities
Ranked first in evaluations for Chinese, English, Japanese, and Korean among models with parameters below 20B.
Technical Advantages
Top Team
More than a hundred top algorithm scientists from global tech giants such as Facebook, Yahoo, Baidu, and more.
Algorithm Mastery
Our technical roadmap encompasses DNN, attention, Bert, LLM, ASR, TTS, NLP, tracking the industry's technological evolution comprehensively.
Scene Understanding
Adapted for applications in over a thousand enterprises.
Application Refinement
Extensive experience in refining applications for a user base of 2 billion globally.
Data Accumulation
Accumulated real user query data in the tens of billions and token data in several tens of trillion over nearly 7 years.
Highly Effective Enterprise-Application LLM
Orion-14B series fine-tuned large language models:
professional scenario capabilities, state-of-the-art ten-billion-parameter models
professional scenario capabilities, state-of-the-art ten-billion-parameter models
General dialogue fine-tuning
Among open-source multilingual large language models below 20B, the best-performing general dialogue model.
Plugin fine-tuning
Enhanced capabilities in Agent, ReAct, and Prompting, delivering results close to a hundred billion parameter models.
RAG fine-tuning
Knowledge boundary control, precision in answers, achieving effects similar to trillion-parameter models.
Long Token fine-tuning
Supports tokens of up to 320K in length, the best among open-source models in token support.
Knowledge extraction fine-tuning
Transforms unstructured data into structured data.
Question-Answer Pair Generation fine-tuning
Generates question-answer pairs while ensuring comprehensive knowledge coverage.
Japanese and Korean fine-tuning
Optimal performance in Japanese and Korean languages among open-source models below 20B.
Orion-14B Large Language Model
Affordable LLM for Enterprise Applications
Suitable for Enterprise Use
After INT4 quantization, the model size is reduced by 70%, inference speed is increased by 30%, with less than 1% performance loss
Run on Affordable Graphics Cards
Such as NVIDIA RTX 3060, capable of achieving 31 tokens per second, approximately 100 characters
Affordable LLM for Enterprise Applications
Suitable for Enterprise Use
After INT4 quantization, the model size is reduced by 70%, inference speed is increased by 30%, with less than 1% performance loss
Run on Affordable Graphics Cards
Such as NVIDIA RTX 3060, capable of achieving 31 tokens per second, approximately 100 characters
Reliable LLM for Enterprise Applications
Private Deployment
Large model servers can be situated within the enterprise intranet, ensuring all data does not connect to the public internet.
Free, Open Source, and Commercially Usable with Community Technical Support
Available for download from communities such as Hugging Face and GitHub.