You type a question into the chat screen and hit the send button. Sometimes the reply starts immediately, but other times ...

Native HIP text-to-image generation with MOSS-Image-SFT from OpenMOSS: a 42-block single-stream diffusion transformer (about 5B parameters), a Qwen3.5 9B text encoder and the FLUX.2 small VAE decoder.

Cartesia Research demonstrates that Hierarchical Networks (H-Nets) learning dynamic tokenization from raw bytes achieve up to ...

This repository is a conversion of nvidia/nemotron-3.5-asr-streaming-0.6b (streaming speech recognition in 40 language-locales) to LiteRT, made with litert-torch 0.10.0 and ai-edge-quantizer 0.10.0 ...

Foundry Robotics raised a $100M oversubscribed Series A led by Marlinspike Partners to build an AI-powered autonomous ...

Phonon-2.5 by Fermion Research, exported to ONNX for ONNX Runtime, including onnxruntime-web on WebGPU in the browser. Phonon-2.5 is a 25-language speech recognizer derived from NVIDIA's Parakeet TDT ...

SoftBank demonstrated up to a 60.5% improvement in 5G uplink L1 throughput under interference using a Transformer-based neural receiver integrated into NVIDIA Aerial. OTA testing with a commercial ...

GTFormer proposes replacing Transformer self-attention with a hierarchical graph neural network, achieving zero-shot length ...

The wider team's NLU design combines a transformer encoder with BiLSTM slot refinement and CRF decoding. Its dialogue model uses an attention-based encoder–decoder. The integrated app loads both ...

A single Transformer-based decoder trained on the OpenBMI dataset classified eight EEG classes spanning motor imagery, P300, ...

Xiaomi-Lens-0 shipped six days ago with pixel masks as tokens. A Transformers PR landed today. What the vendor numbers show, ...

A systematic review finds deep learning exceeds 97 percent accuracy on remote sensing benchmarks but struggles with ...