Alibaba's Qwen 2.5 is the perfect open source model in mathematics and coding

Alibaba has released greater than 100 open-source AI models, including Qwen 2.5 72B, which beats other open-source models in math and coding benchmarks.

Much of the AI ​​industry's attention in open source models has been focused on Meta's efforts with Llama 3, but Alibaba's Qwen 2.5 has significantly closed the gap. The newly released Qwen 2.5 family of models spans 0.5 to 72 billion parameters and includes each generalized baseline models and models focused on very specific tasks.

Alibaba says these models have “enhanced knowledge and stronger skills in math and programming,” with the specialized models focused on programming, math and multiple modalities including speech, audio and vision.

Alibaba Cloud has also announced an upgrade to its proprietary flagship model Qwen-Max, which has not been released as open source. The Qwen 2.5 Max benchmarks look good, but it surely is the Qwen 2.5 72B model that has generated probably the most excitement amongst open source fans.

The benchmarks show that Qwen 2.5 72B beats Meta's much larger flagship model Llama 3.1 405B in some ways, especially in math and coding. The gap between open-source models and proprietary models like those from OpenAI and Google can also be closing quickly.

First users of Qwen 2.5 72B show that the model is just behind Sonnet 3.5 and even OpenAI's O1 models in coding.

Open Source Qwen 2.5 beats o1 models at coding 🤯🤯

Qwen 2.5 achieves higher results than the o1 models when coding on Livebench AI

Qwen is just under Sonnet 3.5 and for an open source mode that's incredible!!

o1 is sweet at some hard-coding problems, but terrible at code completion problems and … pic.twitter.com/iazam61eP9

— Bindu Reddy (@bindureddy) 20 September 2024

Alibaba says these recent models were all trained on its massive dataset, which incorporates as much as 18 trillion tokens. The Qwen 2.5 models have a context window of as much as 128,000 and might generate outputs of as much as 8,000 tokens.

The move to smaller, more powerful, and free, open source models will likely have an even bigger impact for a lot of users than more advanced models like o1. The edge and on-device features of those models mean you may get loads done with a free model in your laptop.

The smaller model, Qwen 2.5, offers GPT-4 encoding at a fraction of the price, and even free if you’ve gotten a good laptop to run it locally.

We have GPT-4 for programming at home! I checkedOpenAI?ref_src=twsrc%5Etfw”>@OpenAI GPT-4 0613 results for various benchmarks and compared them with @Alibaba_Qwen 2.5 7B coder. 👀

> 15 months after the discharge of GPT-0613, now we have an open LLM on Apache 2.0 that works just as well. 🤯

> GPT-4 prices… pic.twitter.com/2szw5kwTe5

— Philipp Schmid (@_philschmid) 22 September 2024

In addition to the LLMs, Alibaba has released a big update to its Vision Language model with the launch of Qwen2-VL. Qwen2-VL can understand videos over 20 minutes long and supports video-based question-and-answer.

It is designed for integration into mobile phones, cars and robots to enable the automation of operations that require visual understanding.

Alibaba also introduced a brand new text-to-video model as a part of its image generator, the Tongyi Wanxiang large model family. Tongyi Wanxiang AI Video can create cinema-quality video content and 3D animations with various artistic styles based on text prompts.

The demos look impressive and the tool is free to make use of, but you have to a Chinese mobile number to register here. Sora will face serious competition when or if OpenAI finally releases it.