Alibaba releases Qwen3.8-27B open weights built to run on a single consumer GPU

Alibaba’s Qwen team released open weights for Qwen3.8-27B under an Apache 2.0 license on Aug. 14, a 27-billion-parameter multimodal model built to run on about 24GB of GPU memory.

The target matters: 24GB is the capacity of a high-end consumer graphics card such as Nvidia’s RTX 4090, so the model is meant to run on a single desktop rather than a data-center cluster. It handles text, image and video input and carries a native 262,144-token context window that Alibaba says extends to 1 million tokens using a technique called YaRN.

Qwen3.8-27B is the compact member of the broader Qwen3.8 line, whose largest model, Qwen3.8-Max, carries roughly 2.4 trillion total parameters with about 95 billion active. The team published the weights on Hugging Face, including an FP8 quantized version that lowers the memory needed to run it.

Alibaba says the 27B model outperforms its own larger Qwen3.7-Plus overall and posts state-of-the-art results among open models on agentic coding, computer use, browser tasks and long-horizon professional work. Those are vendor-supplied claims; the company has not released third-party evaluations, and open-model leaderboards shift week to week.

The release continues Alibaba’s push to put capable open weights within reach of individual developers rather than only cloud customers. An Apache 2.0 license permits commercial use and modification with few restrictions, which tends to speed adoption.

Whether independent benchmarks back Alibaba’s claim that a 27-billion-parameter model beats far larger systems will decide how much the release reshapes what developers run locally.