Alibaba’s Qwen team open-sourced Qwen 3.8-27B on Aug. 14 and released it under the Apache 2.0 license, making the multimodal model available for commercial use.
A 27.8 billion-parameter multimodal model built for local deployment
According to the official model card, Qwen 3.8-27B has about 27.8 billion parameters. It comes with a native context window of 262,144 tokens that can be extended to 1 million, and it supports understanding across text, images and video.
The article said a model at this scale can run on a high-end laptop or a Mac after quantization. That lines up with the recent focus on local AI, where users put a free, offline-capable model directly on their own devices.
Vendor-reported benchmarks show wins in some coding and agent tasks
In the comparison published in Qwen’s official model card, Qwen 3.8-27B outscored Anthropic’s Claude Opus 4.6 Max on several programming and agent benchmarks:
- SWE-bench Pro: 61.7 versus 53.4
- LiveCodeBench v6: 90.3 versus 88.8
- OSWorld-Verified: 84.3 versus 72.7
- Terminal Bench 2.1: 73.0 versus 78.2, where Qwen trailed
The article also flagged a limitation in that comparison. Claude Opus 4.6 Max used officially published scores, while Qwen was evaluated through the Claude Code testing framework, so the test conditions were not identical. The figures are still self-reported by the vendor and have not yet been confirmed by third-party validation.
Clear gains over Qwen3.6-27B
Compared with the previous-generation Qwen3.6-27B, the new release showed marked improvement in agent and coding capability. In the figures cited by the article, OSWorld-Verified rose from 63.9 to 84.3, while DeepSWE increased from 13.3 to 42.2.
ABMedia said the launch continues the recent trend of open-source models moving closer to the front edge of closed-source systems at lower cost. The piece also noted that Chain News recently reported on DeepSeek V4 Pro positioning itself against Claude on price-performance.
As free, locally deployable models become more competitive on some tasks, the article said they could keep pressure on the pricing and defensive moat of OpenAI and Anthropic.