Alibaba Cloud has unveiled the Qwen 2.5 Artificial Intelligence model. The standout feature of this version is its training on data up to 18 trillion tokens, supporting 29 languages including Thai, and allowing nearly unrestricted usage across all generations.
The model supports input of 128K tokens and can respond with 8K tokens, except for the 3B version which can handle input of 32K tokens only. The largest model, Qwen2.5-72B, has shown test results surpassing Llama3.1-70B in almost all tests approaching Llama3.1-405B. However, the 72B version comes with a Qwen License that limits usage to services with less than 100 million users per month.
For smaller-sized models, the Qwen team boasts that any model scoring above 65 MMLU points continuously improves over time. Qwen2.5-3B is the smallest model to exceed 65 points, while the 3B version is distributed for research purposes only.
The Qwen2.5-Coder-7B special edition has outperformed several large models in almost all tests, only losing to CodeStral-22B in the BigCodeBench evaluation.
Qwen2.5 can be downloaded for use with both vllm and ollama platforms, with both platforms supporting tool calling features.
TLDR:
Alibaba Cloud introduces Qwen 2.5 AI model with massive data training, multilingual support, and various model versions catering to different needs and research purposes.
Leave a Comment