The Hub is quieter than usual. Three items stand out. H Company put a computer-use VLM family on the blog. SeaWolf added a self-improving 180B MoE. AutoTrust shipped an open decision model that answers in one forward pass.
On the Hub
H Company Holo4
H Company blogged Holo4 on Sep 28. The family is a generalist computer-use VLM with a 27B CC BY-NC 4.0 model and a 35B-A3B Apache-2.0 model. The cards have been on the Hub since Sep 24.
The models run GUI, code, MCP and API tasks from one weights file. On OSWorld 2.0 the 27B scores 61.7 percent and the 35B-A3B scores 30.9 percent. The blog lists BF16, FP8, NVFP4 and Q4 GGUF formats.1
Darwin-180B-RSI
SeaWolf added Darwin-180B-RSI to the Hub on Sep 28. The card is under Qwen Community 1.0. It is a 180B MoE vision-language model built from Qwen3.8-Flash-Next with recursive self-improvement and zero-token confidence. The README lists author-claimed AIME 2026 100 percent as majority vote over 16 samples, verified false.23
autotrust/JEV-27B
AutoTrust blogged JEV-27B on Sep 27. The model is Apache-2.0 and built on Qwen3.8-27B. It adds a frozen backbone with a detachable LoRA and a 24-slot decision head. It returns calibrated probabilities in one forward pass with no decoding. The card notes mean KL 0.0186 on 25,376 held-out questions.45
