Qwen3-Coder-30B-A3B-Instruct

Quality drops 9.8 points to 86.5 as category mix shifts from coding

Qwen3-Coder-30B-A3B-Instruct experienced a notable quality decline this window, falling from 96.3 to 86.5 overall. The most significant change is a shift in tested categories, with coding tests absent from the current window while new categories emerged. Multilingual performance remains the model's strongest area, maintaining exceptional scores at 100 compared to 99 previously. Creative work held relatively steady, moving from 90 to 88. However, the newly tested reasoning category scored 75, and factual performance came in at 83, both pulling the overall average down. The absence of coding tests is particularly notable given this model's specialized positioning and its perfect 100 coding score in the previous window. On the positive side, latency improved by 16 percent, dropping from 4655ms to 3913ms at median, making the model more responsive for interactive use cases. With only 5 test runs in each window, these results should be considered preliminary. Users should note that while the model continues to excel at multilingual tasks and maintains decent creative capabilities, the current test mix suggests more variability in reasoning and factual domains than previously observed.