Hugging Face
AI Platform
Open-source ML platform, model hub, Transformers library.
Watched repos
Recent activity
-
The article discusses the creation of an interactive Halloween projection, "MonsterHouse", using NVIDIA's Nemotron Voice Agent and various models. The system uses NVIDIA's Nemotron Speech ASR to transcribe visitor speech, Bolt E4B to generate responses, Magpie TTS Zeroshot to synthesize speech, and Audio2Face-3D to animate the character's mouth. The system is built on a single workstation and runs on NVIDIA's RTX PRO 6000 and RTX PRO 4000 GPUs. AI summary
Read more → -
IBM has released Granite Switch 4.1.1, a patch update that brings incremental improvements to the modular language model architecture, including an updated Granite Speech ASR model for multilingual speech-to-text across 10 languages and an improved Decision adapter. This update is fully backwards compatible with version 4.1 and can be composed and deployed without recomputing the base model. The update demonstrates the speed and efficiency of IBM's model update process. AI summary
Read more → -
AnySpeechRecognizer is an open model for speech recognition on Apple platforms, providing a choice of recognition engines, including NVIDIA's Parakeet and Alibaba's Qwen3-ASR, with speaker diarization capabilities. The package mirrors Apple's Speech framework and uses Core ML models on the Hugging Face Hub. AnyDecisionModel is an open model for decision-making, using a trained decision head like Cloudflare's Clef, and supports multiple backends, including MLX, Core AI, and llama.cpp. AI summary
Read more → -
Mistral Large 4, a multimodal, multilingual, Mixture-of-Experts (MoE) model, takes text and images as input and generates text, with 1T total parameters, 49B active parameters, and a context window of 1M tokens. AI summary
Read more → -
Project ProtoStar aims to make publicly funded scientific research more usable by releasing datasets, models, and code in AI-ready formats, supported by open infrastructure and practical publishing guidance, with the goal of increasing the scientific value of the work released. The initiative is hosted by Hugging Face and brings together universities, research organizations, industry, and the open science community to support the release of scientific artifacts in machine-actionable formats. AI summary
Read more → -
NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model that transcribes 40 languages, translates 31 of them, and attributes speakers in 13, with top leaderboard results for transcription, speaker attribution, and translation. The model can be fine-tuned to adapt to domain vocabulary and accents, with a case study demonstrating a 45.8% relative reduction in word error rate on held-out speakers. The model can be fine-tuned using the NeMo Speech fine-tuning tools and Granary v2 data pipeline. AI summary
Read more → -
LFM3 is a new generation of Liquid Foundation Models (LFMs) that combines text, vision, audio, omni, and decision models onto one shared backbone, allowing for specialized intelligence in various modalities. LFM3 models can be fine-tuned on a single GPU for specific tasks, enabling deployment on devices, private infrastructure, or the cloud. Six specialized models (text, vision, audio, omni, and decision) are built on top of the shared LFM3 backbone, each with its own unique architecture and training recipe. AI summary
Read more → -
Large language models (LLMs) can produce highly variable judgments of novel research ideas, with small changes in the evaluation setup resulting in dramatically different verdicts. This instability is particularly pronounced when evaluating AI-generated ideas, where the choice of prompt, reasoning budget, and model can significantly impact the outcome, and more search or reasoning does not necessarily produce more consistent judgments. AI summary
Read more → -
Researchers developed LitPivot, a system that helps researchers navigate literature-initiated pivots by organizing literature around a proposal being edited, offering critiques supported by excerpts from selected studies, and suggesting revisions. The system was rated as better grounded in the literature than proposals developed with a text editor and an AI chat tool, and participants reported a stronger understanding of the surrounding literature and greater self-reported improvements in their ideas' novelty and usefulness. LitPivot can support both early exploration and the refinement of an existing research direction. AI summary
Read more → -
Hugging Face provides a Space tool to check which models run on AWS Trainium, with the Space listing the number of Trainium cores the model is sharded over, compilation time, and peak device memory. The results are stored in a public dataset, allowing developers to query the data directly. To run a model on Trainium, users can select the model, view the run.py script, and install the necessary packages on a Trainium2 instance. AI summary
Read more → -
Researchers at Microsoft and Hugging Face have developed a method to turn database grading into a reward signal for post-training agents, using the ThinkingBox framework. By using the executable checks from the ThinkingBox database as rewards, the agents were able to improve their performance on tasks outside the ThinkingBox benchmark, with average success rates increasing from 51.70% to 60.78% and tasks completed correctly on all twenty attempts rising from 38 to 107. AI summary
Read more → -
OlmoEarth v1.3 is a state-of-the-art satellite embedding model that generates 128-dimensional vectors for every 10-meter pixel on Earth, producing embeddings for the entire planet. The model reads radar and multispectral satellite time series data, and its embeddings have been precomputed and shared with the community for free, allowing users to fine-tune small classifiers on top of the embeddings without requiring GPU access or significant computational resources. AI summary
Read more → -
IBM's Granite time-series models can be used to create ensemble forecasts, combining the predictions of four complementary architectures (PatchTST-FM-r1, PatchTST-FM-r2, Patched Conformer, and FlowState-r1.1) into one predictive distribution. The ensemble can be created using either the Probabilistic Ensemble, which pools quantile forecasts with equal influence, or the IQR Ensemble, which weights members according to their forecast spread. Both methods can be used to improve forecast accuracy and robustness without requiring developers to select the strongest individual architecture in advance. AI summary
Read more → -
NVIDIA Kumo Relational is a pre-trained relational foundation model that outperforms supervised machine learning on relational data without requiring task-specific training or feature engineering. It operates directly on enterprise databases, respecting their tables, foreign keys, and timestamps, and achieves state-of-the-art performance on 41 benchmarks across 15 relational databases. Kumo Relational can be used as a GPU-native structured-data-models library or as an NVIDIA NIM, and its Predictive Query Language (PQL) allows for simple and flexible querying of relational data. AI summary
Read more → -
Specializing the vocabulary of large language models (LLMs) can significantly reduce inference costs by up to 77% in real-world applications. By fine-tuning a general LLM on a domain-specific dataset, a new method called VocabX adapts the vocabulary to only include tokens that are relevant to the task at hand, eliminating unnecessary tokens and reducing computational costs. This approach can lead to substantial savings in inference time and energy, making it an attractive solution for applications where efficiency is crucial. AI summary
Read more → -
Here is a summary of the article in 2-3 plain sentences for a developer/AI-ML audience: To profile a PyTorch model on AWS Trainium using TorchNeuron, you can use the NeuronConfig object to control the data streams and NeuronCores to capture, and the NeuronProfiler object to export the framework trace and system traces. By analyzing the key_averages() table, you can identify which operations consume Neuron device time and compare their Self CPU time, and use Neuron Explorer to visualize the trace and follow an operation through framework dispatch, runtime scheduling, and hardware execution. The article also shows how to use torch.compile to fuse operations into a single kernel dispatch, reducing device time by up to 6x. AI summary
Read more → -
The IBM Granite Speech 5.0 collection includes three updated speech models: granite-speech-5.0-2b, granite-speech-5.0-470m-turboctc, and granite-speech-4.1-2b-plus, which offer improved accuracy, robustness, and control in transcription, translation, and speaker attribution tasks. granite-speech-5.0-2b is the successor to granite-speech-4.1-2b and supports multilingual transcription and translation, while granite-speech-5.0-470m-turboctc is a multilingual version of the original Turbo CTC model, offering faster and more robust transcription. granite-speech-4.1-2b-plus is updated for enterprise use cases, such as meetings and multi-speaker recordings, with improved speaker attribution, timestamps, and keyword biasing. AI summary
Read more → -
NVIDIA Cosmos3-H-Dreams utilizes a 16-billion-parameter model pretrained on Open-H-Embodiment to generate interactive, real-time clinical surgical scenes with low latency generation, addressing the challenges of visual realism and fidelity in surgical simulation. AI summary
Read more → -
Researchers have introduced FairStress, a benchmark for measuring demographic fairness in large language models (LLMs) beyond the final answer. The benchmark assesses models' behavior in 300 realistic decision scenarios, including ambiguous cases where no clear winner emerges, and evaluates their responses to adversarial prefixes that aim to sway their decisions. By analyzing FairStress results, developers can diagnose hidden asymmetries and test proposed fixes to develop more consistent and trustworthy LLMs. AI summary
Read more → -
Learned memory improves a basic ReAct agent with GPT-4.1 to the top of the τ²-bench leaderboard on customer service tasks without updating model weights, outperforming frontier models in both retail and airline domains. The agent achieves higher accuracy and reliability by analyzing its training-task attempts and distilling successes and failures into short behavioral rules, which are then added to the agent's prompt. AI summary
Read more →