Alibaba on Friday released the open weights of Qwen3.8-27B, a 27-billion-parameter multimodal dense model, on Hugging Face and ModelScope under the Apache 2.0 licence, honoring the commitment it made at the August 3 launch of its 2.4-trillion-parameter Qwen3.8-Max flagship. The release lands one day after a Hugging Face state-of-open-models report, cited by Bloomberg, showed Qwen accumulating more than 3 billion global downloads over the past six months, ahead of Google’s 418 million and Meta’s 227 million over 2026.

The 27B model is a serious piece of engineering. It pairs a Hybrid Gated DeltaNet and Gated Attention architecture with a 262,000-token native context window extensible to 1 million via YaRN, and accepts text, image, and video inputs. The Qwen team reports SGLang running the model at 206.1 tokens per second on a single Nvidia RTX 5090, with vLLM also supported. Benchmark scores land at 61.7 on SWE-Bench Pro, 73.0 on Terminal-Bench 2.1, and 90.3 on LiveCodeBench v6. Unlike the larger Qwen3.8-Max open-weight drop, the 27B ships with non-thinking mode and image inputs, and outperforms Qwen3.7-Plus overall.

The strategic frame matters more than the numbers. Per an emailed statement, Alibaba has now open-sourced more than 460 models, spawning over 300,000 ecosystem derivatives. That’s the volume playbook Meta ran with Llama in 2023, executed against a Chinese cost base and with faster cadence.

The catch sits one tier up. Reuters reported on August 7, citing two people familiar with the plans, that Alibaba intends to require large commercial users to share a portion of revenue from the Qwen3.8-Max weights. Apache 2.0 at 27B, tollbooth at the flagship: distribution as loss leader, monetization on the frontier.

Sources