Alibaba drops 100+ open-source Qwen 2.5 models and a new video generator

Alibaba Cloud released over 100 open-source Qwen 2.5 models at its annual Apsara Conference on September 19, offering a full range from lightweight 0.5B variants up to full-scale 72B parameter architectures.

What it can do

Qwen 2.5 delivers significantly upgraded reasoning in mathematics and coding alongside support for more than 29 languages. The collection spans multiple modalities including text, audio, and vision, complete with dedicated code and math specialist checkpoints.

Alongside the text models, Alibaba introduced Qwen2-VL, an upgraded vision-language model capable of comprehending videos over 20 minutes long and responding to video-based questions. It is engineered for low-latency reasoning on mobile phones, vehicles, and robotic hardware.

Alibaba also debuted a new text-to-video generator under its Tongyi Wanxiang brand. Built on a Diffusion Transformer (DiT) architecture, it turns Chinese or English prompts into realistic or 3D animated video clips and converts static images into motion video.

For software engineers, the company rolled out AI Developer, an assistant designed to automate requirement analysis, code generation, and bug fixing. Alibaba also refreshed its proprietary flagship Qwen-Max (Qwen2.5-Max), claiming parity with premier frontier models in language and reasoning tasks.

How to get and run it

The release covers base models, instruct-tuned versions, and pre-quantized variants across several bit precisions. That means developers can skip manual quantization when deploying to memory-constrained hardware.

The smallest 0.5B checkpoints run easily on local consumer devices, while the flagship 72B version is geared toward heavier cloud or multi-GPU environments.

All open checkpoints are live on Hugging Face and Alibaba's ModelScope platform for immediate download. Alibaba noted that the Qwen family has already crossed 40 million total downloads, fueling over 50,000 derivative community models.

Why it matters

If you build local AI workflows or domain-specific coding agents, having a turnkey family from edge-ready sizes up to 72B with official quantization saves massive setup overhead. You get state-of-the-art math and coding capabilities without being tethered to closed APIs.

This release also raises the baseline for the entire open-weights space, putting direct competitive pressure on Meta's Llama lineup. We think releasing the complete model suite at once makes Qwen one of the most practical and flexible open-source foundations currently available.