The Hub moved on September 11. A music model, an Olympiad math recipe, a new Nex-mini GGUF, and a compressed ASR from XPENG topped the day.
On the Hub
YuE2-3B
m-a-p updated YuE2-3B. The card adds WildSongBench numbers and a runtime note for RTX 4090. The model is text-to-audio under CC-BY-NC-4.0. It claims 6.9632 SongBench average best-of-8 versus Suno v5 at 6.8721. Generation is 71.04 s with 11.18 GiB peak VRAM on a 24 GB RTX 4090. Weights are about 6.79 GiB plus a 507 MiB VAE. No GGUF is offered. 1
NVIDIA IMO Gold
NVIDIA posted an open recipe for IMO Gold. The paper is on Hugging Face Daily Papers and arXiv. It describes a natural language proof pipeline with no formal prover or tools. It claims 30 out of 42 at IMO 2026. Two specialist checkpoints were updated on the Hub under OpenMDW-1.1. Both are about 1.12 TB and not consumer runnable. 2
bartowski Nex-N2.5-mini-GGUF
bartowski published the first imatrix llama.cpp pack for Nex-N2.5-mini. The GGUF card is Apache-2.0 and built with llama.cpp b10896. The recommended Q4_K_M file is 22,318,512,256 bytes, about 22.32 GB. It includes mmproj and per tensor layout maps. The size fits one RTX 5090 32GB with headroom. The base model is Qwen3_5MoeForConditionalGeneration. 3
X-AuT XPENG
XPENG-AI released X-AuT weights and a paper. The work prunes Qwen3-ASR-0.6B from 18 to 14 layers. The audio tower shrinks from 186.376M to 147.794M under CC-BY-NC-4.0. The card claims encoder latency down 21.4 percent on XPENG in vehicle PPU and 11.4 percent on H800. The checkpoint is about 743.84M total and 1.68 GiB used storage. No GGUF is offered. 4
