Hugging Face

Hugging Face announced Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs.

Hugging Face has launched a centralized leaderboard specifically for benchmarking Text-to-Speech (TTS) and voice cloning models. The platform provides standardized evaluation metrics for multilingual speech synthesis performance.

NVIDIA has released Kumo Tabular, a specialized machine learning framework designed to optimize tabular data prediction tasks. The model architecture achieves state-of-the-art performance by balancing predictive accuracy with computational efficiency.

Hugging Face has introduced a source-aware verification framework designed to integrate with Model Context Protocol (MCP) agents. This system enables agents to cross-reference generated claims against verified source documents to reduce hallucinations.

Hugging Face updated its Inference Endpoints to support automatic scale-to-zero serverless deployment for any open model on the Hub.

Hugging Face has released Holo4, a specialized model architecture designed to enable generalist computer-use agents. The model provides native capabilities for interpreting and interacting with graphical user interfaces across diverse operating systems.

Hugging Face announced Accelerating vision-language models with LFM2.5-VL-DSpark.

Hugging Face has integrated NVIDIA Warp and MjWarp into its robotics ecosystem to enable high-performance GPU-accelerated simulation. This integration allows developers to execute physics kernels directly in Python while maintaining C++ performance levels for robotics learning workflows.

Hugging Face announced **Know Who Spoke When: Build Real-Time, Multi-Speaker AI with NVIDIA Nemotron 3 Diarization**.

The Hugging Face Transformers library now natively supports loading and running GGUF-formatted quantized models via the llama.cpp backend. This integration allows users to execute compressed models directly within the Transformers ecosystem without requiring external conversion tools.

Hugging Face has integrated the UK AISI's Inspect framework with the EvalEval platform to standardize model evaluation protocols. This integration enables automated, reproducible execution of safety and capability benchmarks for large language models.

Hugging Face has hired Jun Kim, the creator of the oMLX library, to formalize support for the MLX framework within its ecosystem. This move signals a strategic commitment to optimizing Apple Silicon-based machine learning workflows.

Hugging Face introduced a novel model pruning methodology that treats transformer block removal as an Ising optimization problem. This approach utilizes physical system modeling to identify and remove redundant layers while minimizing performance degradation.

Hugging Face has released version 1.0 of the 'tokenizers' library, marking a transition to a stable API. This release focuses on performance optimizations for encoding and decoding processes and establishes long-term API stability.

Hugging Face announced Your Agent Aced the Task. Will It Do It Again?.

Hugging Face has introduced an asynchronous implementation of Group Relative Policy Optimization (GRPO) that supports LoRA fine-tuning across distributed HF Jobs. The architecture eliminates the requirement for NCCL (NVIDIA Collective Communications Library) by utilizing an S3-compatible bucket and a proxy for state synchronization.

Hugging Face has integrated the AUTOMATIC1111 Stable Diffusion web UI into a native Gradio-based workflow. This transition replaces legacy interface components with Gradio's modular blocks to improve UI responsiveness and component extensibility.

IBM has released the Granite Time Series PatchTST-FM-r2 model, a foundation model specifically optimized for time-series forecasting. The model is distributed under the Apache 2.0 license, enabling unrestricted commercial use.

Hugging Face has introduced a refined safety filtering methodology that enables models to distinguish between harmful and benign sub-topics within a broader category. This approach replaces blanket topic refusals with granular classification to improve model utility while maintaining safety guardrails.

Hugging Face announced NeoMME: an efficient Multimodal-native and Multilingual Encoder.

Hugging Face announced Give Your Coding Agents a Memory You Own.

Hugging Face announced Training a coding model to paint watercolours with TRL and OpenEnv.

Hugging Face announced Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps.

Hugging Face announced Real-Time Intelligence with IBM Time Series Models on Confluent.

Hugging Face announced BenchMIRT: What are LLM benchmarks actually measuring?.