The Hub is quiet this morning. Four items set the day. Qwen adds an eight step turbo of Image 2.1. TII writes up a 1.6B Arabic speech model. Hugging Face tags a Python only Tokenizers patch. AllenAI describes a new GPU cluster scheduler.
On the Hub
Qwen-Image-2.1-Turbo
Qwen put Qwen-Image-2.1-Turbo on the Hub on October 9. It is a 7B eight step version of Qwen-Image-2.1. The pipeline is QwenImage21Pipeline with CFG set to 1 and prefix KV cache. License is qwen-research. Community builds followed, including an Unsloth FP8 version. This is not the September Image 2.1 release.12
Falcon ASR
TII published a blog on October 7 about Falcon ASR. The model is 1.6B and focuses on Arabic with an Emirati dialect emphasis. It also handles English, French, Spanish and Portuguese on the same weights. The blog claims WER of 20.92 percent on Arabic, 22.73 percent on Emirati, and 5.74 percent mean on English. Word level timestamps are included. A demo Space is available.34
Tokenizers v0.23.3
Hugging Face tagged tokenizers v0.23.3 on October 9. It is a Python only patch. The release loosens huggingface_hub to allow v2, adds pyright hints, and fixes a truncation and overflowing break from 0.23.1. Other language packages are unchanged.5
AllenAI Impactful scheduling
AllenAI published a blog on October 9 about a new GPU cluster scheduler. It replaces a priority based system with GPU time budgets, hierarchical fair share and time slicing. The post describes thousands of H100, B200 and B300 GPUs across clusters of 88 to 1024 GPUs for about 150 researchers, with two to three times overcommit.6
