Sakhanda Wire
NVDA $154.00 -1.37% MSFT $506.28 -2.26% GOOGL $280.87 -1.00% META $620.34 -0.97% AMZN $256.88 -0.34%

Hugging Face

111 stories · all sources

Hugging Face

Microduck Walking Policies Move from Isaac Lab Simulation to Real Robots

Pollen Robotics and Hugging Face have released a workflow for training the open-source Microduck bipedal robot in simulation and transferring the resulting walking policy to physical hardware. The tutorial uses NVIDIA Isaac Lab 3.0 and Newton to model servo response, friction, delays and gearbox backlash, factors that can affect performance outside simulation.

Hugging Face

VocabX cuts measured inference cost for an essay grader by 77%

A vocabulary-specialization method called VocabX reduced the measured inference cost of a Portuguese essay grader by 77.0%, according to results from a paper that is still under review. The system uses Qwen3-4B, but its original tokenizer was designed for many languages and domains, leaving it with a 151,669-entry vocabulary and fragmented output for the grader’s repeated JSON structures.

Hugging Face

TorchNeuron profiling compares three RMSNorm implementations on AWS Trainium

A profiling guide published on October 21, 2026, examines how to analyze inference performance with TorchNeuron, the native open source PyTorch backend for AWS Trainium. The first part assumes no previous profiling experience and describes how to configure torch.profiler with the Neuron backend, capture a trace and interpret device and CPU timings.

Hugging Face

IBM updates Granite Speech models for accuracy, speed and speaker attribution

IBM has released three updates to its Granite Speech collection: the new granite-speech-5.0-2b and revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The models target multilingual transcription and translation, fast edge or on-device processing, and recordings involving multiple speakers.

Hugging Face

NVIDIA introduces Kumo Forecast for context-enhanced time series

NVIDIA Kumo Forecast, formerly called NV-Tesseract Forecasting, is a multivariate time-series model that combines pretrained representations with patterns retrieved from historical data. Its backbone is the MOMENT-1-large Transformer, with an approximately 340-million-parameter encoder and a trainable forecasting head. The direct mode predicts future values from recent observations, while Domain-Aware Representation and Retrieval (DARR) adds related historical patterns found through nearest-neighbor retrieval.

Hugging Face

SimReady Central offers over 2,000 assets for robotics simulation

SimReady Central, a new library on Hugging Face, brings together more than 2,000 OpenUSD assets prepared for simulation workflows. The free, ungated collection is intended as a starting point for robotics training environments and AI factory digital twins. Its catalog includes robot arms and grippers, warehouse equipment, racks, cooling systems and other infrastructure.

Hugging Face

NVIDIA introduces a real-time generative simulator for surgical procedures

NVIDIA has introduced Cosmos3-H-Dreams, a research simulator designed to generate interactive video of clinical surgery. The system was trained on synchronized video and robot-kinematics recordings from cholecystectomy procedures performed with the CMR Surgical Versius platform. NVIDIA describes it as the first real-time generative surgical simulator trained on clinical cholecystectomy recordings to support live kinematic control.

Hugging Face

FairStress tests hidden demographic asymmetries in LLM decisions

AI2 researchers have introduced FairStress, a benchmark designed to assess demographic fairness in large language models beyond their final choices. The benchmark covers 300 two-candidate decision scenarios across 29 domains, including hiring, lending, healthcare, education, housing, and criminal justice. It varies whether demographic identity is absent, implied, coded through names, or explicitly stated, and tests models with matched persuasive prompts called adversarial prefixes.

Hugging Face

Learned memory lifts GPT-4.1 to the top of τ²-bench retail

ALTK-Evolve uses GPT-4.1 to extract behavioral guidelines from an agent’s previous task attempts and adds them to the system prompt. The model weights are not changed, and no human annotation is used. On τ²-bench retail, the resulting ReAct agent achieved a 67.5% Pass^4 score, meaning it completed all four trials for that share of tasks and exceeded every retail entry in the October 12, 2026 leaderboard snapshot.

Hugging Face

Hugging Face adds native TPU support through TorchTPU

Hugging Face has introduced TorchTPU, a native PyTorch backend that treats Google’s Tensor Processing Units as another device. The company says the integration enables Transformers, Diffusers and Kernels to run directly on TPUs, allowing users to load a Hub model and assign it to “tpu” without moving their code to JAX or relying on torch_xla.

Hugging Face

AnyLanguageModel 1.0 brings unified model access to Swift apps

AnyLanguageModel 1.0 has been released after a year of development and dozens of releases. The Swift package is designed as a drop-in replacement for Apple’s Foundation Models framework, allowing developers to keep the same session code while choosing system, local or hosted models. The release accompanies swift-huggingface 1.0 and swift-xet 1.0.

Hugging Face

PyTorch 2.15 expands MPS capabilities on Apple silicon

PyTorch 2.15 adds major MPS improvements for Apple silicon Macs. The update introduces support for torch.float8_e4m3fn tensors and scaled matrix multiplication, allowing FP8 values to use half the storage of BF16. It also requires macOS 15 or later and Python 3.11 or later.