Károly Zsolnai-Fehér presents Qwen 3.8 Max as an open model competing with leading closed systems while charging substantially lower API prices. The model supports multimodal inputs, a one-million-token context window and agentic workflows intended to continue working without constant supervision.
Demonstrations include a long-running coding task that starts from an empty folder and iteratively writes, tests and repairs its own implementation over 16 days. The model is also shown reproducing and improving research work, and building websites and applications.
The full model is too large for most personal hardware, but Károly Zsolnai-Fehér expects smaller versions to make the broader release useful for local deployment. He points to earlier compact Qwen models as strong daily drivers and treats the coming family as the practical opportunity for users with modest compute.
Károly Zsolnai-Fehér also highlights the model's result on Humanity's Last Exam as evidence of rapid open-model progress. He argues that strong capability and low prices could force other providers to reduce costs, while the promised weight release would extend the work to open research and local experimentation.
Watch on YouTube
