Alibaba open-sources Qwen3.8-27B multimodal model with 262K context, runs on consumer GPUs
Alibaba Qwen open-sourced Qwen3.8-27B native multimodal dense model on August 14 evening. The 27B parameter model supports text/image/video input, Apache 2.0 license, runs locally on RTX 4090 with ~17GB VRAM after quantization.
The Alibaba Qwen team officially open-sourced the Qwen3.8 model family on the evening of August 14. Qwen3.8-27B is a 27-billion-parameter native multimodal dense model that supports text, image, and video input. It has a native context length of 262K tokens and can extend to 1M via YaRN. The model uses the Apache 2.0 license, and after FP8 quantization fits into approximately 17GB VRAM, runnable on RTX 4090-class consumer GPUs.
On benchmarks, Qwen3.8-27B's DeepSWE coding agent score jumped from 13.3 (Qwen3.6) to 42.2, surpassing Opus 4.6 Max on some agent benchmarks. Terminal-Bench 2.1 climbed from 63.4 to 73.0. The model pairs with the flagship Qwen3.8-2.4T-A95B (adapted via FlagOS to 9 chips including T-Head, NVIDIA, Moore Threads, and Huawei Ascend) to form a high-low product matrix covering enterprise deployment and local development scenarios.
Alibaba has open-sourced more than 460 models in total. The Hugging Face "State of Open Models" report released August 16 put Qwen's August Hub downloads at 2.045 billion, with cumulative global downloads exceeding 3 billion, surpassing Google (418 million) and Meta (227 million), and over 300,000 derivative models. Qwen is now one of the largest open-model families globally.