Alibaba open-sources Qwen3.8-27B multimodal, tops Hugging Face trending on Aug 17

Alibaba's Tongyi Qianwen team officially open-sourced the Qwen3.8-27B large language model on August 14, 2026. The model is a native multimodal dense model with about 27 billion parameters, released under the Apache 2.0 license. The official model card reports SWE-bench Pro at 61.7, exceeding Opus 4.6 Max's 53.4; DeepSWE 1.1 improved from the prior 27B model's 13.3 to 42.2.

The model natively supports a 262K context window, outputs across image, text, and video modalities, and provides controllable reasoning depth and tool calling. By August 17, Qwen3.8-27B had topped the Hugging Face global model trending list. Mainstream inference frameworks vLLM, SGLang, and Ollama have completed adaptation.

On the same day, MediaTek announced Day-0 adaptation of Qwen3.8 on the Dimensity C-X1 automotive cockpit platform and Dimensity flagship mobile chips. Counterpoint Research noted that on-device compute is becoming the next focal point of the AI industry, with local deployment offering fast response and stronger security.