NVIDIA and Hugging Face Detail Nemotron 3 Nano Benchmarking
A new NVIDIA-authored Hugging Face blog lays out an open evaluation recipe for Nemotron 3 Nano using the NeMo Evaluator framework.
AI & ML
Frontier models, research results, agents, and developer tooling.
AssetOpsBench Aims to Test AI Agents on Real-World Industrial Tasks
A new benchmark from IBM Research, hosted on Hugging Face, attempts to close the divide between synthetic AI agent tests and the demands of industrial asset operations.
Dynamic Surface Codes Give Quantum Error Correction a Time-Varying Twist
Google researchers describe dynamic surface codes, a time-varying approach that could expand the toolkit for quantum error correction.
Hugging Face Introduces Waypoint-1, a Real-Time Interactive Video Diffusion Model
The model from Overworld aims to bring real-time interactivity to diffusion-based video generation.
Hard-Braking Events Could Flag High-Risk Road Segments
Google Research suggests sudden braking telemetry can identify dangerous road stretches before crash records accumulate.
Falcon-H1-Arabic: TII’s Hybrid Model for Arabic AI
A new open Arabic language model from the Technology Innovation Institute combines dense and mixture-of-experts layers to improve efficiency and performance.
Smartwatch Gait Metrics Promise Continuous Health Tracking
Google Research shows how smartwatch sensors can estimate advanced walking metrics, potentially enabling passive health monitoring outside the lab.
Microsoft's Diff Transformer V2 Sharpens Attention to Cut Noise
A new attention mechanism from Microsoft improves language model efficiency by focusing on differences between attention heads rather than their raw values.
Hugging Face Overhauls Tokenizers in Transformers v5
The new tokenizer design in Transformers v5 aims to simplify usage, improve clarity, and make the system more modular.
NVIDIA Cosmos Reason 2 Targets Reasoning for Physical AI
A new NVIDIA model aims to bring advanced reasoning to physical AI systems, according to a Hugging Face announcement.
AprielGuard Brings an Adversarial Robustness Layer to LLMs
ServiceNow AI's AprielGuard is a dedicated guardrail for keeping LLM systems safe against harmful content and adversarial inputs.
DeepMind updates Gemini audio models for voice experiences
Google DeepMind says improved Gemini audio models are aimed at making voice interactions more powerful and natural.
DeepMind unveils Gemini 3 Flash for fast, low-cost AI
Google DeepMind's new model aims to deliver frontier-level intelligence with faster performance and lower cost.
Google Research's 2025 Review: Bolder Breakthroughs, Bigger Impact
Google's year-in-review post signals that 2025 was defined by more ambitious research and a push toward wider real-world impact.
NVIDIA Pairs DGX Spark with Reachy Mini Robot for Embodied AI
A Hugging Face blog post highlights NVIDIA's effort to combine its DGX Spark platform with the Reachy Mini robot for developing embodied AI agents.
ML Writer's Year in Review: Career, Health, and Travel
A personal retrospective from an ML-focused writer frames 2025 around health, career, and travel rather than AI milestones.
Gemini Tool Gives Automated Paper Feedback at STOC 2026
Google Research has described a Gemini-backed assistant that provides automated feedback on theoretical computer science papers at STOC 2026.
Decomposition Lets Small Models Beat Bigger Ones at Intent Extraction
Google Research reports that breaking intent extraction into smaller subtasks lets compact models outperform larger ones.