Sakhanda Wire
NVDA $153.77 -1.52% MSFT $504.91 -2.53% GOOGL $281.52 -0.78% META $620.99 -0.87% AMZN $256.80 -0.37%

Hugging Face

111 stories · all sources

Hugging Face

ToyBlocks offers a controlled testbed for robot learning

Ai2 has introduced ToyBlocks, an open simulation framework designed to study robot learning in a deliberately limited setting. It models a robot arm working with blocks on a tabletop, using realistic physics and appearance while supporting many tasks, including pushing, flipping, sorting, searching and stacking. Images and letters on block faces can require the robot to reveal hidden sides before acting.

Hugging Face

Egyptian and Nigerian English added to Open ASR Leaderboard

The Open ASR Leaderboard has added two private evaluation splits for Egyptian and Nigerian English through a collaboration with Liva AI. The Egyptian split contains 1.34 hours of conversational speech, while the Nigerian split contains 1.55 hours. Both datasets consist of spontaneous conversations and provide punctuated, cased transcriptions that retain disfluencies.

Hugging Face

NVIDIA releases four open Cosmos 3 models for world simulation

NVIDIA has added four open-weight checkpoints to its Cosmos 3 family of world models for interactive simulation. The Cosmos3-Nano-Transfer-Auto model generates synchronized driving scenes across up to 11 camera views and can produce joint RGB and LiDAR data. Cosmos3-Nano-Sim-Auto creates successive camera and LiDAR segments from driving controls, supporting streaming and longer interactive rollouts.

Hugging Face

Phonon brings multilingual text-to-speech to local devices

Phonon, a compact text-to-speech model, is designed to generate speech directly on phones, laptops, single-board computers and in web browsers, without a server or network connection. Its developers have released two versions of about 65 MB and 75 MB. Together, they support English, French, German, Spanish and Portuguese, with more than 100 built-in voices. Custom voices can be created in Gradium Studio through voice cloning or voice design.

Hugging Face

Dharma Field Extractor turns documents into structured JSON

Dharma-AI has introduced Dharma Field Extractor, a 2-billion-parameter model designed to pull requested information from document images and return it as structured JSON. It can operate without a predefined schema, identifying fields from labels, values and table layouts, or follow an exact schema with specified names and data types. Missing fields are returned as null, while tables are represented row by row, including nested structures.

Hugging Face

StateBench measures how models track changing information in long documents

Researchers have introduced StateBench, a benchmark for evaluating how language models follow changing information about entities across long documents. The tasks reflect situations such as variables being reassigned in code, characters moving in a story, or objects changing hands. They range from finding an unchanging or most recent value to following cumulative updates and dependencies between variables.

Hugging Face

Palette trains a 20B model to create decks in a custom design language

Palette is an open-weight system designed to generate presentation decks in a user-defined design language while running on private infrastructure. Its developers fine-tuned a 20B model with LoRA on synthetic data based on the IBM Design Language. The system separates the work into two roles: a designer creates a structured JSON brief covering content, layout and design rationale, while a coder produces pptxgenjs JavaScript for each slide. The slides can be generated in parallel, making the coding stage about 6–7 times faster.

Hugging Face

CUGA outperforms Hermes and Deep Agents in AppWorld comparison

An IBM Research team compared three agent harnesses using the same open-weight model, GLM-5.3-flash, on 90 AppWorld tasks covering 30 scenarios. The benchmark spans nine applications and 457 APIs, including email, payments and file storage. CUGA completed 73 tasks, or 81.1%, while Hermes completed 42 tasks, or 46.7%, and Deep Agents completed 27, or 30.0%. CUGA also passed all tasks in 20 scenarios, compared with eight for Hermes and two for Deep Agents.

Hugging Face

CWE-bench v2 expands public testing of cybersecurity models

CWE-bench v2 introduces a public benchmark of 100 tasks for evaluating AI agents on software security work. Each task uses a real open-source repository, a short instruction and a goal: identify and fix a weakness while preserving existing functionality. Scoring is programmatic and deterministic, allowing results to be reproduced without additional cost. The tasks cover 91 CWE categories, and a separate held-out set is used to track performance on unseen problems.

Hugging Face

Transformers proposes contracts for model integration testing

Transformers is testing a proposed standard called Model Integration Contracts (MICs) to detect regressions before they reach the main codebase. The initiative followed a bug introduced on July 30, 2026, when enabling PyTorch scaled dot-product attention for T5 allowed the UMT5 decoder to access future tokens during training. The issue was fixed on September 28 after a small CPU fixture reproduced it.

Hugging Face

NVIDIA Nemotron powers an interactive Halloween haunted house

NVIDIA employee Rick Grandy created MonsterHouse, an interactive Halloween projection that listens to trick-or-treaters and replies as a grumpy old pirate ship. The garage-door character becomes more accommodating when visitors say “please” and “thank you,” and can distribute treats as part of the performance.

Hugging Face

Granite Switch 4.1.1 adds multilingual speech and decision updates

Granite Switch 4.1.1 is a patch release for the modular language-model architecture introduced with version 4.1. The update is fully backward compatible with 4.1 and does not require migration, according to the article published on October 29, 2026.