Alibaba’s Qwen team has released Qwen 3.8 27B, a 27-billion-parameter open-source vision-language model available under the Apache 2.0 license on Hugging Face. The model, announced August 14, represents the most capable generation in the Qwen open-model family to date and marks the first time a Qwen model of this class ships with native image and video understanding built in.
Qwen 3.8 27B features a 262,000-token native context window, extendable to 1 million tokens through YaRN interpolation. The architecture uses a Gated DeltaNet and Gated Attention block layout across 64 layers with a hidden dimension of 5,120. This hybrid design allows the model to handle hour-scale video understanding alongside traditional text and image tasks without requiring separate modal pipelines.
Benchmarks Show Competitive Performance
On established evaluation suites, Qwen 3.8 27B scores 61.7 on Scale AI’s SWE-Bench Pro, 73.0 on Terminal-Bench 2.1, and 90.3 on LiveCodeBench v6. The model also achieves 42.2 on MMLU and 89.2 on GPQA Diamond. These numbers position it competitively against both larger proprietary models and other open-weight alternatives at similar parameter counts.
The model also supports flexible thinking control, allowing developers to toggle between standard and extended reasoning modes depending on task complexity. This is a significant addition for agentic workflows where multi-step reasoning and tool use are critical.
Part of a Broader Qwen 3.8 Rollout
The 27B dense model follows the release of Qwen 3.8-Max on August 3, a 2.4-trillion-parameter mixture-of-experts flagship available only through Alibaba Cloud’s API. While Qwen 3.8-Max targets enterprise deployments requiring maximum capability, the 27B variant is designed for local self-hosting on consumer and developer hardware, including single GPUs.
Alibaba has also released a Qwen 3.8-27B-FP8 quantized variant for further reduced memory requirements. The company says integrations with popular frameworks including vLLM, SGLang, Transformers, and llama.cpp are available at launch.
The release continues Alibaba’s strategy of maintaining a dual-track approach: proprietary flagships for cloud revenue and permissive open-source models for ecosystem building. Previous Qwen open-weight models, particularly Qwen 3.6-35B-A3B, have seen widespread community adoption for self-hosted deployments in enterprise and research settings.
With this release, the 27B class of open-source models now includes a strong multimodal contender that can process images, video, and long documents on a single GPU, lowering the barrier for developers building local AI agents and coding assistants.
Sources: Hugging Face model card; Alibaba Qwen blog; AI Weekly; LLM Stats
discussion