Hugging Face Daily Digest Oct 24, 2026

Tencent's WeVisOne-VL leads Daily Papers, with Apache-2.0 2B and 4B weights that have been on the Hub since 22 October. Hugging Face published a Trainium lookup so you can check which Transformers models actually run unmodified. Unsloth refreshed its MiniMax-M3.1-Flash GGUF, and the parent model is now number 14 on the default trending page. Daily Papers moved off yesterday's Deep Search and FAPL lead to WeVisOne.

On the Hub

Tencent WeVisOne-VL

Tencent's WeVisOne-VL (WeChat Vision All-in-One) is Daily Papers number one. The paper, Unified Visual Perception from Objects to Pixels, covers open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation, and promptable alpha matting. Official 2B and 4B cards have been on the Hub since 22 October, Apache-2.0, built on Qwen3-VL.¹²³

The 2B card is 9.817 GB (9.767 GB of safetensors plus a 38.1 MB pixel decoder). The 4B card is 19.373 GB. Install the GitHub package and import WeVisOneVL. Tencent marks it for research, not for identifying people or high-impact autonomous decisions, and the pixel decoder rejects video grids. Code and a project page are public.⁴⁵

Trainium model-support lookup

Hugging Face posted "Does Your Model Run on AWS Trainium? Look It Up", a changelog that points at a Space and a Parquet dataset. The Space lists which popular Hub Transformers models run unmodified on Trainium. It emits a run.py that loads with from_pretrained, uses a static cache, and compiles with torch.compile(backend="neuron"). Tests cover text-generation and image-text-to-text for models that fit on four Trainium2 devices (16 cores). A PASS is not a throughput claim.⁶⁷

This follows the TorchNeuron backend post from 20 October. The results dataset last changed today. Compiled graphs sync with hf buckets sync hf://buckets/aws-neuron/torch-neuronx-neff-cache/neuronxcc-<version>.⁸⁹

Unsloth MiniMax-M3.1-Flash GGUF

Unsloth updated MiniMax-M3.1-Flash-GGUF today. The parent MiniMaxAI/MiniMax-M3.1-Flash, posted 17 October, is now number 14 on the default trending page. MiniMax describes a multimodal mixture-of-experts for coding, agents, computer use, and cybersecurity. It takes text, image, and video, with a 1 million token context and about 428 billion total parameters (23 billion active per token).¹⁰¹¹¹²

The Unsloth pack is still titled "Work in progress", under the minimax-community license. The smallest official quant is UD-Q3_K_XL at 193.951 GB. The GGUF card itself appeared 23 October. Unsloth also hosts a BF16 mirror of the parent.

Daily Papers flip to WeVisOne

Daily Papers moved off yesterday's Deep Search, FAPL, and Deep Fusion Memory lead. Tencent's WeVisOne-VL is now number one at 227.¹³¹⁴

FAPL (Making MLLM Latents Matter) is at 165, still the 22 October paper, with no Hub model card. Deep Search in Real-World Web Environments is at 152, arXiv 19 October, GitHub OpenDS, no Hub card. Deep Fusion Memory sits at 40, arXiv 18 October, with no GitHub and no Hub card.¹⁵¹⁶¹⁷