On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability

Benchmark Model Rank Results
visual-question-answering-vqa-on-realworldqaQwen3.8-Flash-Next#1Accuracy: 88.5
computer-use-agents-on-androidworldQwen3.8-Flash-Next#2Success Rate: 84.5
code-generation-on-multipl-eQwen3.8-Flash-Next-Base#4Accuracy: 79.09
code-generation-on-swe-bench-multilingualQwen3.8-Flash-Next (mini-SWE-agent)#7Accuracy: 81.0
code-generation-on-swe-bench-proQwen3.8-Flash-Next#7Resolved: 62.5
multi-task-language-understanding-on-mmmluQwen3.8-Flash-Next-Base#7Accuracy: 84.86
multiple-choice-question-answering-mcqa-on-gpqa-diamondQwen3.8-Flash-Next#7Accuracy: 91.7
code-generation-on-nl2repoQwen3.8-Flash-Next#12Accuracy: 48.1
multi-task-language-understanding-on-mmlu-reduxQwen3.8-Flash-Next-Base#15Accuracy: 90.68
multiple-choice-question-answering-mcqa-on-supergpqaQwen3.8-Flash-Next-Base#17Accuracy: 51.36
code-generation-on-deepsweQwen3.8-Flash-Next (mini-SWE-agent)#18Pass@1: 0.587
math-word-problem-solving-on-mathQwen3.8-Flash-Next-Base#25Accuracy: 72.78
code-generation-on-livecodebench-v6Qwen3.8-Flash-Next#62Pass@1: 0.919
multi-task-language-understanding-on-mmlu-proQwen3.8-Flash-Next-Base#63Accuracy: 73.23