Just now, Alibaba Qwen3.8-27B is open source! Home graphics card can run and can view pictures and videos

Zhidongxi Author｜Jiang Yu Editor｜Xinyuan Zhidongxi reported on August 14 that just now, Alibaba Qianwen officially open sourced the Qwen3.8-27B model weights. Qwen3.8-27B is a 27 billion parameter native multi-modal dense model that supports image and video understanding. The native context length reaches 262K and can be expanded to 1 million Tokens through YaRN. Compared with the ultra-large-scale MoE model, the 27B Dense architecture has a lower deployment threshold. After quantification, it can be run on home-grade graphics cards, making it more suitable for local deployment and lightweight application development. The model is open using the Apache 2.0 protocol, and developers, scientific research institutions and enterprises can download, deploy and commercially use it for free. ▲Qwen3.8-27B open source homepage (Source: Hugging Face) This is the latest open model of the Qwen3.8 series. Previously, Qianwen has opened the weight of the Qwen3.8-2.4T-A95B model, which is also the Max-level base on which the Qwen3.8-Max released on August 3 is based. Compared with the flagship version with 2.4 trillion total parameters and 95 billion activation parameters, the Qwen3.8-27B released this time further compresses the Agent, multi-modal and long context capabilities of Qwen3.8 into a 27 billion parameter Dense model, focusing more on local and lightweight deployment. According to the test results, Qwen3.8-27B has significantly improved compared to the previous generation Qwen3.6-27B in tasks such as programming, long-term office work, and Computer Use. Many results exceeded Qwen3.7-Plus, which has a larger parameter scale; in some benchmarks such as SWE-bench Pro and OSWorld-Verified, its results also exceeded Claude Opus 4.6 Max. After the model was open sourced, it quickly attracted the attention of overseas developers. Some netizens commented that the current performance of Qwen3.8-27B is "impressive", especially when there are only 27 billion parameters.