Hugging Face

Hugging Face — AI Infrastructure

Changelogs

NVIDIA releases Nemotron-Nano-9B-v2-Japanese, tops sub-10B category on Nejumi Leaderboard//NVIDIA has released a Japanese-optimized version of its Nemotron-Nano-9B-v2 model, achieving state-of-the-art performance in the under-10B parameter category on Japan's Nejumi Leaderboard 4. The model combines advanced Japanese language understanding with proven agent capabilities in a lightweight, deployable package designed for enterprise on-premises deployment.

releasemodelfeature

Hugging Face releases CUDA kernel skill for Claude and Codex agents//Hugging Face has released an agent skill that enables AI coding agents like Claude and Codex to automatically generate production-ready CUDA kernels for deep learning models. The skill packages domain expertise about GPU architecture, PyTorch integration patterns, and optimization strategies into a reusable module that agents can apply to real transformers and diffusers pipelines.

featuresdkintegrationplatform

Transformers.js v4 Preview Debuts on NPM with New WebGPU Runtime and 10x Build Speed Gains//Transformers.js v4 preview is now available on NPM under the `@next` tag, bringing a complete rewrite with a new WebGPU runtime and major performance improvements. The release includes support for ~200 model architectures, cross-runtime compatibility (browsers, Node, Bun, Deno), and architectural optimizations that deliver 4x speedups for embedding models and 10x faster builds.

releasefeaturesdkperformanceopen-source

SyGra Studio v2.0.0 debuts visual no-code editor for synthetic data generation workflows//ServiceNow has released SyGra Studio, a visual interface for building and executing synthetic data generation pipelines without writing YAML or scripts. The platform enables drag-and-drop workflow composition with live execution monitoring, inline debugging, and real-time cost tracking.

releasefeatureplatformopen-source

NVIDIA releases Nemotron ColEmbed V2 multimodal models; 8B variant ranks #1 on ViDoRe V3 benchmark//NVIDIA has released the Nemotron ColEmbed V2 family, a set of late-interaction multimodal embedding models in 3B, 4B, and 8B sizes designed for visual document retrieval. The flagship 8B model achieves 63.42 NDCG@10 on the ViDoRe V3 benchmark, ranking first in its class for enterprise document retrieval tasks.

releasemodelfeature

Hugging Face launches community-driven model evaluations with decentralized benchmarking//Hugging Face is introducing a decentralized evaluation system that allows the community to submit and verify model benchmark results directly on the Hub. The system links benchmark datasets, model repositories, and evaluation results through git-based pull requests, creating transparent and reproducible eval scores across the community.

featureplatformapiopen-source

H Company releases Holo2-235B-A22B, reaching 78.5% accuracy on UI localization benchmarks//H Company has released the Holo2-235B-A22B Preview model, their largest UI localization model to date, achieving state-of-the-art performance on Screenspot-Pro and OSWorld G benchmarks. The model introduces "agentic localization," an iterative refinement approach that delivers 10-20% relative accuracy gains by allowing the model to progressively improve its predictions across multiple steps.

releasemodelfeature