The Hub is quiet on official posts. The day brings a MoE with a new quant, an Intel export update, and a community architecture note. No new weights from the big labs today.
On the Hub
Xing4.0-29B-A4B
China Telecom's XingChen-AGI put Xing4.0-29B-A4B on the Hub on September 16.1 The card describes a 29B total, 4B active MoE with 40 layers and 256K context, trained on Ascend NPU with MindSpore. The model is Apache-2.0 and lists Transformers, vLLM, SGLang and KTransformers as serving options. The GGUF repo was touched this morning and the IQ4_NL file is about 20 GB.2 Card-claimed benchmarks include SWE-bench Verified 75.00 versus Qwen3.6-35B-A3B 76.00.
Optimum-Intel v2.2.0
Hugging Face's community blog posted Optimum-Intel v2.2 on September 18.3 GitHub tagged v2.2.0 on September 17.4 The release pairs optimum-intel 2.2.0 with OpenVINO 2026.4, OpenVINO GenAI 2026.4 and NNCF 3.4. The notes list Mistral 3, DeepSeek-OCR-2, Muse Glimmer 30B, Ouro, Fun-ASR and Qwen-Image support. Gemma 4 and Unified now accept video input in this stack.
Layer-Feedback Transformer
Banaxi-Tech published a community post on Layer-Feedback Transformer on September 19.5 The idea reuses adjacent layers to get more compute per parameter with zero new weights. For N unique layers the author claims 3N minus 4 block executions. In a controlled 10M parameter setup with 500M tokens, the author claims LFT reaches 36.86 percent on Base Bench versus 32.57 percent for standard. No weights are on the Hub and the result is author-claimed and not compute-matched.
