Sakhanda Wire
NVDA $154.51 -1.04% MSFT $506.49 -2.22% GOOGL $281.27 -0.86% META $621.51 -0.78% AMZN $256.90 -0.33%

Hugging Face

111 stories · all sources

Hugging Face

Mistral Large 4 joins Hugging Face with multimodal capabilities

Mistral Large 4, also known as Le Chonk, has been released on Hugging Face as a multimodal and multilingual Mixture-of-Experts model. It accepts text and images and generates text, with about 1T total parameters, 49B active parameters and a context window of 1M tokens. Official checkpoints are available in FP8 and NVFP4, alongside GGUF versions from Unsloth and the llama.cpp team. The model uses the Mistral Ecosystem Open-Weight License, an MIT derivative with limits on model-as-a-service offerings.

Hugging Face

Project ProtoStar seeks to make publicly funded science AI-ready

Project ProtoStar will bring together universities, research organizations, industry and the open science community to make publicly funded scientific datasets, models and code easier to discover, reuse and combine. The initiative supports releasing these resources in AI-ready, machine-actionable formats on open infrastructure. Its founding partners include Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University, The George Washington University, the Astera Institute, Hugging Face and the Open Research Community Accelerator.

Hugging Face

AnySpeechRecognizer and AnyDecisionModel bring open models to Apple platforms

Two new Swift packages, AnySpeechRecognizer and AnyDecisionModel, extend model choices for speech recognition and structured decisions on Apple platforms. AnySpeechRecognizer mirrors Apple’s SpeechAnalyzer-related API while allowing developers to select another recognition engine. It supports NVIDIA’s Parakeet and Alibaba’s Qwen3-ASR running locally through Core ML, alongside Apple’s default engine.

Hugging Face

NVIDIA releases multilingual Nemotron 3.5 Transcribe model

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model available through Hugging Face and ModelScope. The single checkpoint supports transcription, speaker attribution, translation and audio question answering. It transcribes 40 languages, translates 31 of them to or from English, identifies speakers in 13 languages, handles recordings of up to one hour and can label as many as eight speakers.

Hugging Face

LFM3 brings specialized multimodal models onto one backbone

LiquidAI has announced LFM3, a new generation of Liquid Foundation Models designed as a base for specialized applications. The family covers text, vision, audio, omni and decision tasks, with six models built from the LFM3-1.7B-Base backbone. Developers can fine-tune each model on a single GPU and deploy it on the device where an application runs.

Hugging Face

Study finds LLM judgments of research novelty are highly unstable

A new study from Ai2 examines how reliably large language models can determine whether scientific research ideas are novel. The researchers released the Old Ideas, Novel Problems benchmark, data and code, along with a pipeline for creating additional evaluation sets. The benchmark tests both absolute novelty ratings and comparisons between two ideas.

Hugging Face

LitPivot links research-idea revision to relevant literature

Ai2 and collaborators at the University of Pennsylvania have developed LitPivot, a system for shaping early-stage research ideas through ongoing comparison with academic literature. The tool uses papers retrieved by Asta Paper Finder and organizes them around the part of a proposal being edited: the problem, proposed contribution, or evaluation plan.

Hugging Face

New Space Checks Transformers Models on AWS Trainium

Hugging Face has introduced the torch-neuronx model support Space, which checks whether selected Transformers models can run without modification on AWS Trainium. The Space initially shows models chosen from the Hub’s most-downloaded and trending entries, while separate views follow those rankings. Users can search for a model and see its result, the number of Trainium cores used, total runtime including compilation, and peak device memory. Compiler and Transformers versions can also be selected.

Hugging Face

ThinkingBox Turns Database Checks into Rewards for Agent Training

Microsoft and Hugging Face report that ThinkingBox-Training is now available on GitHub. The system evaluates an agent by examining the final state of a database rather than relying only on the tool calls or the agent’s claim that a task is complete. Its executable checks can produce a binary reward for reinforcement learning.

Hugging Face

OlmoEarth v1.3 offers global satellite embeddings and an exploration tool

OlmoEarth v1.3 is a satellite embedding model designed to turn radar and multispectral time series into reusable features. It processes annual Sentinel-1, Sentinel-2 and Landsat data, producing a 128-dimensional vector for each 10-meter pixel on Earth. The first 64 dimensions can also be used as a smaller embedding. Precomputed embeddings for the planet, in both sizes, are being shared with the community.

Hugging Face

IBM Granite ensemble combines four models for time-series forecasting

IBM has introduced Granite Timeseries Ensemble, a forecasting recipe that combines four pre-trained time-series models through a common interface. The approach produces a median point forecast and quantile predictions without retraining the models or fitting ensemble weights. Its default configuration includes Granite PatchTST-FM-r1, Granite PatchTST-FM-r2, Granite FlowState-r1.1 and Granite TTM-r3, representing patched Transformer, patched Conformer, state-space and MLP-mixer architectures.

Hugging Face

Kumo Relational reports stronger results without task-specific training

NVIDIA has introduced Kumo Relational, a pre-trained foundation model designed to make predictions directly from relational databases. The model uses a predictive query and up to 10,000 historical examples as context to estimate outcomes such as customer churn, demand and fraud. It does not require task-specific training, feature engineering or flattening tables, although fine-tuning remains available for maximum accuracy.