The Hub is quiet on official posts. Activity centers on a week-old image model touched this morning. Two community posts on answer verification from September 20 share a test set and disagree on who is on top.
On the Hub
Qwen-Image-2.1
Qwen put Qwen-Image-2.1 on the Hub on September 14 and the card was last modified this morning.1 The official Qwen blog is dated September 20.2 The model is a unified text to image and image editing DiT with native RGBA output. The visual generation component is 7B parameters with 32 Single Stream DiT layers. The license is Qwen Research, not Apache.
Zero Token Confidence
FINAL Bench published a community article on September 20 about reading hidden states instead of generating tokens.3 The probe is a 46 KB npz file on Darwin 397B, a card created in June.4 Author claimed AUC is 0.7394 for ZTC versus 0.5000 for self reported confidence. Author claimed a surface baseline of length and formatting scores 0.7036. The probe needs no training data.
JEV ecosystem benchmark
mayafree published a community article on September 20 comparing 13 verifiers on one test set.5 Author claimed the surface baseline of length and formatting scores 0.7036 and beats eight of the 13 systems. Author claimed ZTC on Darwin 397B scores 0.7364 and JEV scores 0.7350. The gap is 0.0014 with a 95 percent interval that includes zero, so the authors treat them as tied. Both posts use the same test set.
