An open model you can download is winning at agent work
Alibaba published Qwen3.8-27B's weights on Hugging Face on 14th Aug’26 under Apache 2.0, free to use commercially. Its model card reports 84.3 on computer-use tasks against 72.7 for Opus 4.6 Max, and 61.7 on SWE-bench Pro against 53.4. Close to 2.9 million downloads in a month, with 839 quantised builds for Ollama, LM Studio and llama.cpp.
What it actually means
Vendor benchmarks flatter the vendor, so read those numbers as a claim, not a result. The claim still matters. Running your own model is now a real option, not a compromise. Test it on one live task and compare honestly. The question is no longer whether open models are good enough. It's what you'd still pay a frontier lab for.
The question is no longer whether open models are good enough. It's what you'd still pay a frontier lab for.
2.9Mdownloads in a month of a model you can run yourself
Source: Hugging Face (Qwen model card) · Wednesday 26 August 2026