AI & Software

A new open speech-to-speech pipeline combines Cerebras fast inference, Gemma 4, and modular components for natural, low-latency voice interactions.

Microsoft Foundry offers a curated catalog of Hugging Face models deployable on Managed Compute with enterprise security, governance, and observability.

Mount Hub repos directly into any cloud job with no egress fees. Use hf:// URLs and HF_TOKEN to run on 20+ clouds.

IBM Research launches ScarfBench, an open benchmark evaluating AI agents on cross-framework Java migration, revealing configuration and runtime dependencies as top hurdles.

From optimization theory to biology and markets, specialization emerges under finite resources. Fit beats breadth.

Hugging Face has revamped its Kernels project with a new repository type, improved security, revamped CLIs, broader framework support, and foundations for agentic kernel development.

NVIDIA and Hugging Face team up to bring distributed diffusion training to any Diffusers-format model on the Hub, with no checkpoint conversion and full support for LoRA and full fine-tuning.

The EvalEval Coalition and Hugging Face integrate structured evaluation results into model pages, making scores traceable and comparable.

LeRobot v0.6.0 introduces world model policies that imagine the future, new VLAs, unified reward models, six new benchmarks, and faster data loading.