alibaba recently officially unveiled the next-generation flagship model in the qwen series—qwen3.6-max-preview. as an early preview version of the series, users can now engage in interactive conversations with it on qwenstudio and invoke it via the alibaba cloud bailian api under the name “qwen3.6-max-preview.” compared with the previous generation, qwen3.6-plus, the new model delivers significant improvements across multiple dimensions, including agent-based programming, world knowledge, and instruction following.
according to alibaba, qwen3.6-max-preview achieved the highest scores in six major programming benchmark tests, covering swe-benchpro, terminal-bench2.0, skillsbench, qwenclawbench, qwenwebbench, and scicode. specifically, its agent-based programming capabilities have improved markedly: scores on skillsbench increased by 9.9 points, on scicode by 10.8 points, and on nl2repo and terminal-bench2.0 by 5.0 and 3.8 points, respectively. these results represent a significant step forward in the model’s programming capabilities.
in addition, qwen3.6-max-preview has also seen a notable enhancement in its grasp of world knowledge, with supergpqa scores rising by 2.3 points and qwenchinesebench improving by 5.3 points. its instruction-following ability is equally impressive, as toolcallformatifbench scores increased by 2.8 points. taken together, these improvements position the new model as a major innovation in the field of intelligent programming.
although currently available as a preview version, alibaba states that the model is still under active development, with future iterations set to undergo continuous optimization in order to deliver even more powerful features and performance gains.
