Mercury 2.5 Is Fast, but Got Only 13 of 43 Rare Javanese Words Right
We tested two models released on 8 and 10 September 2026 on four tasks from three Indonesian and Javanese benchmarks. Mercury 2.5, a diffusion model from Inception, answered with a median latency of about 1 second on all four tasks, but got only 13 of 43 rare Javanese words right, far below GPT-5.6 Luna, which it is claimed to match. DeepSeek V4.1 Flash answered 58 of 60 slang items and 37 of 38 speech-level items correctly, no real difference from DeepSeek V4 Flash, at 6.7 to 10.8 times its predecessor's cost.
1,041 msMercury latency13/43Rare words10.8×V4.1 cost