Publisher
Stories with reporting from Hugging Face Blog, ranked by what is gaining ground now. Each story links to the original articles.
Cloudflare announced Clef, a set of open-weight models intended to select among predefined options rather than generate text. The release includes 9B- and 27B-parameter models and a platform for adapting them to decision tasks.
NVIDIA describes fine-tuning Nemotron for gold-level results in both the International Olympiad in Informatics and the International Mathematical Olympiad. The post appears on Hugging Face.
A technical timeline examines a July 2026 incident involving an intrusion at a frontier AI lab.
NVIDIA presents Cosmos-H-Dreams, an effort to bring real-time generative simulation to surgical robotics.
The Technology Innovation Institute introduces Falcon ASR.
Multiverse Computing presents Quantization-Aware Healing, a compressed 4-bit model that it says outperforms its full-precision original.
Ai2 introduces OlmoEarth embeddings, which can be exported from OlmoEarth Studio for downstream analysis.
The Open ASR Leaderboard adds its first language from the Global South, extending its evaluation of open automatic speech recognition systems.
A Hugging Face project reports that Transformers now runs llama.cpp quantized models.
Jun Kim, creator and maintainer of oMLX, joins Hugging Face to support the MLX community. The move brings the project’s creator into a role focused on that community.
Hugging Face introduces @huggingface/kernels, a collection of more than 200 WebGPU kernels for running AI workloads locally.
Liquid AI presents LFM2.5-DSpark as delivering up to 3.2x faster inference. The post appears on Hugging Face.
Ai2 open-sources AstaBrief, the report-generation model used in Asta.
ServiceNow AI presents AutoSynthData, a system for generating training data for enterprise agents.
NVIDIA presents Kumo Tabular, describing it as a model that advances the accuracy-efficiency tradeoff for tabular prediction.
A project examines tokenizers v1 through measurements of encoding, decoding, and scaling. Its stated focus is open-source and open-science work in artificial intelligence.
A project fine-tunes a 350M model for structured outputs in 100 GRPO steps. Its stated aim is to advance open-source and open-science work in artificial intelligence.
A team reports lessons from reproducing 2,200 papers from ICML. The work is framed as part of an effort to advance AI through open source and open science.
The Technology Innovation Institute presents Falcon-Emirati, a language model focused on Emirati dialect, culture, and nuance. The post appears on Hugging Face.
UK AISI and EvalEval are working to make benchmark results reproducible. The effort is framed around open source and open science.
