Alibaba’s Alibaba Group-owned Tongyi Lab released Qwen3.8-27B on August 14, a 27.8-billion-parameter open-weight model built to run on a single high-end GPU while matching results usually reserved for far larger systems. The dense, multimodal model reads text, images and hour-scale video, ships under an Apache 2.0 license, and natively handles a 262,144-token context window that can be stretched toward one million tokens.
What changed
Compared with the prior Qwen3.6-27B release, the new model posts sharp gains on tasks that stress real-world coding and reasoning. Qwen’s own benchmark data shows SWE-bench Pro rising from 53.5 to 61.7, Terminal-Bench 2.1 climbing from 63.4 to 73.0, and OSWorld-Verified — a test of controlling a full desktop environment — jumping from 63.9 to 84.3. The company attributes the gains to a hybrid attention design that mixes linear “Gated DeltaNet” layers with standard attention layers across 64 total layers, an architecture meant to cut memory use during long tasks rather than simply adding parameters.
Fits on a desktop GPU
Unlike the trillion-parameter flagship models several labs have shipped this year, Qwen3.8-27B is aimed at developers without a data center. Alibaba has also published an FP8-quantized version, and community-made 4-bit builds compress the weights enough to run on a single 24GB consumer card such as an Nvidia RTX 4090, rather than the multi-GPU racks that larger open models require.
The release continues a pattern of labs pushing efficiency alongside scale: Meta’s own 30-billion-parameter Muse Glimmer, released as an open-weight agent model, and Apple’s China-specific model built with Alibaba’s help both point to growing competition in the mid-size open-weight tier that used to be dominated by far larger, closed systems. Qwen3.8-27B adds to Alibaba’s broader Qwen family, which spans from small on-device models to trillion-parameter flagships, and reinforces Alibaba’s push into open-weight AI as a way to build developer mindshare outside China.