Alibaba's Qwen team released open weights for Qwen3.8-27B on August 13-14, a dense 27.78-billion-parameter multimodal model shipped under the permissive Apache 2.0 license. The model accepts text, images, and video, carries a native 262,144-token context window, and is built to run locally on a single GPU, needing roughly 56GB of VRAM at BF16, about 28GB at FP8, or as little as 14 to 16GB at 4-bit quantization before accounting for the KV cache.
The release positions Alibaba directly against a wave of local-first, single-GPU-friendly models from Meta and Google that have shipped in recent weeks. On Alibaba's own reported benchmarks, Qwen3.8-27B beats Meta's similarly sized Muse Glimmer 30B on several agentic tasks, with Terminal-Bench 2.1 rising from 63.4 to 73.0 and DeepSWE 1.1 climbing from 13.3 to 42.2 compared with the prior Qwen3.6-27B generation. On SWE-Bench Pro and CoWorkBench, the open 27B model also outperforms Alibaba's own API-only Qwen3.7 Plus. As with any vendor-reported scores, these numbers have not yet been independently verified across the full suite.
The 27B model arrives roughly a week after Alibaba's 2.4-trillion-parameter Qwen3.8-Max flagship, a sparse mixture-of-experts model with about 95 billion active parameters, launched on August 3 with a priced API at 2 dollars per million input tokens and 6 dollars per million output tokens. Open weights for the Max model landed first, with the smaller 27B trailing by about a week and a half after Alibaba's original one-week target lapsed. The gap between announcement and open release has drawn some scrutiny, though the final Qwen3.8-27B repository is now live on Hugging Face under the official Qwen organization.
