SOURCE ARCHIVE · ORIGINAL LINKS INCLUDED
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech-to-text model that transcribes, translates, and attributes speakers across multiple languages with promptable control.
Hugging Face introduces LFM3, a family of specialized multimodal models designed for on-device deployment, featuring improved agentic capabilities and reduced parameter counts compared to predecessors.
A new Hugging Face study finds large language models often fail to reliably judge the novelty of research ideas, with minor prompt changes causing accuracy swings of over 50 percentage points.
Hugging Face introduces LitPivot, a tool that helps researchers refine early-stage ideas by dynamically linking proposals to relevant literature and suggesting revisions based on retrieved papers.
Hugging Face launches a tool to check which Transformers models run unmodified on AWS Trainium, providing run scripts and performance metrics via a public Space.
Hugging Face introduces ThinkingBox-Training, a method converting database state checks into reinforcement learning rewards to improve agent reliability on unseen tasks.
Hugging Face releases OlmoEarth v1.3, a satellite embedding model with precomputed global embeddings and a browser tool for change detection and classification without technical barriers.
IBM’s open Granite time-series models now support ensemble forecasting via a new recipe that combines four architectures into a single, default probabilistic forecast without retraining or manual tuning.
NVIDIA introduces Kumo Relational, a pre-trained model that predicts outcomes directly from enterprise databases without training or feature engineering, outperforming supervised methods on relational benchmarks.
Hugging Face demonstrates a 77% reduction in inference costs for specialized small language models by optimizing tokenizers for domain-specific tasks rather than general use.
Hugging Face details how to profile PyTorch models on AWS Trainium using TorchNeuron, explaining system-level tracing and performance analysis workflows.
IBM updates its Granite Speech collection with three new and revised speech-language models, enhancing transcription, translation, and speaker attribution across multiple languages and use cases.
NVIDIA introduced Kumo Forecast, a time-series model combining pretrained Transformers with historical pattern retrieval and interpretability tools for improved forecasting accuracy.
NVIDIA introduces Cosmos3-H-Dreams, a generative surgical simulator trained on clinical cholecystectomy videos, enabling real-time interactive control of surgical instruments within realistic in-body scenes.
Hugging Face introduces FairStress, a benchmark to evaluate demographic fairness in large language models beyond final answers, revealing hidden biases in decision robustness and explanations.
ALTK-Evolve, a learned-memory system using GPT-4.1, tops the τ²-bench customer service leaderboard without model updates or frontier models by distilling behavioral rules from training tasks into guidelines appended to...
Apple’s AnyLanguageModel 1.0 simplifies local and hosted model integration for Swift apps, replacing Foundation Models with a unified API and expanding supported providers.
PyTorch 2.15 introduces FP8 support on MPS, faster language-model decoding, and new Metal kernels for audio, video, and training workloads on Apple silicon Macs.
Hugging Face introduces TorchTPU, a native PyTorch backend enabling direct TPU support for Transformers, Diffusers, and Kernels, eliminating prior software barriers.
Hugging Face and partners released a 2.8 trillion-token Indic pretraining corpus with an open GPU pipeline, aiming to expand Indian-language AI resources and technical model development.
Hugging Face introduces Beam, a 501B-parameter open MoE model by Reflection, designed for coding and agentic tasks with a 1M-token context window and Apache 2.0 licensing.
Hugging Face’s Granite Switch 4.1 integrates chat, decision-making, and speech-to-text into a single endpoint, reducing GPU overhead and accelerating multi-task workflows with shared KV caches and optimized adapters.
Hugging Face released Apertus 1.5, enhancing its 8B and 70B models with multimodal capabilities, longer context windows, and improved instruction-following under an open Apache 2.0 license.
Researchers unveiled HiRO-ACEv2, an AI system generating high-resolution climate data at 3 km resolution for extreme weather events at a fraction of the cost of traditional models like X-SHiELD.
Hugging Face introduces LFM2.5-24B-A2B, a 24-billion-parameter reasoning model optimized for on-device agent deployment with only 2 billion active parameters per token, enabling cost-effective, private, and fast local...
Ai2 replaced a priority-based GPU scheduler with a budget-driven system to prioritize impactful research while maintaining full cluster occupancy across thousands of NVIDIA GPUs.
Hugging Face releases two open decision models, d1-3B and d1-omni-600M, optimized for edge devices and benchmarked across seven public datasets.
Hugging Face reports Nemotron models achieved gold-level results in IOI 2026 and IMO 2026 through fine-tuning and specialized inference pipelines, releasing open resources for reproducibility.
Hugging Face introduces Falcon-Emirati-7B, a dialect-specialized model built on Falcon-H1-Arabic to better understand and generate Emirati Arabic with cultural nuance.
Hugging Face introduces ThinkingBox, a benchmark grading AI agents by backend state changes rather than tool calls, revealing widespread consistency failures across 507 workflows and 12 models.