Releases·Tool Release
Transformers 5.18.0 adds Nemotron 3 Diarization, NemotronH Omni, HyperCLOVAX Vision V2, and GTE
Hugging Face released transformers 5.18.0, adding support for four new model architectures: NVIDIA's Nemotron 3 Diarization and NemotronH Omni, NAVER's HyperCLOVAX Vision V2, and Alibaba's GTE encoder. The release lets practitioners run these models through the standard Transformers API. It covers streaming audio diarization, multimodal reasoning over text, images, video, and sound, and long-context multilingual retrieval.
0 engagement1 sourceWed, Sep 30, 2026, 04:46 PM
AI Overview
- Transformers 5.18.0 adds Nemotron 3 Diarization, an open-weight streaming speaker diarization model handling up to eight speakers via AOSC and FIFO queues (#49056)
- Nemotron 3 Diarization's single checkpoint spans latency profiles from an 80 ms input buffer to a 30.4 s offline buffer with 10 ms output frame resolution
- NemotronH Omni pairs NVIDIA's hybrid Mamba-Transformer LM with a RADIO vision encoder and optional Parakeet sound encoder to reason jointly over text, images, video, and sound (#46509)
- HyperCLOVAX Vision V2 combines NAVER's HyperClovaX backbone with a Qwen2.5-VL vision encoder, supporting text, image, video inputs and thinking-token chain-of-thought reasoning (#44314)
- GTE, proposed in the mGTE paper, is a BERT-style encoder using RoPE, gated MLP, and post-residual layer norm, backing Alibaba's gte-*-v1.5 and Snowflake's arctic-embed-m-v2.0 checkpoints
Entities
Related
Products·launched
speech-swift: Nemotron 3 speaker diarization ported to Apple Silicon for local voice agents
speech-swift is a free Apache-2.0 Swift library and CLI that runs NVIDIA's Nemotron 3 speaker diarization locally on Apple Silicon, ported to Core ML INT8 and MLX INT8. Diarization gives voice agents a timeline of which anonymous speaker is active; no LLM is central, though it complements ASR pipelines. It solves the problem of running diarization without cloud services for Mac developers building voice agents.
speech-swift is a free Apache-2.0 Swift library and CLI that runs NVIDIA's Nemotron 3 speaker diarization locally on Apple Silicon, ported to Core ML INT8 and MLX INT8. Diarization gives voice agents a timeline of which anonymous speaker is active; no LLM is central, though it complements ASR pipelines. It solves the problem of running diarization without cloud services for Mac developers building voice agents.
4 engagement1 sourceredditThu, Sep 24, 2026, 09:05 PM
Releases·Model Release
inclusionAI releases Realtime-Venus: 9B full-duplex audio-visual omni model on Hugging Face
inclusionAI published Realtime-Venus on Hugging Face: a full-duplex audio-visual interaction system with two checkpoints, including a 9B Omni model adapted from MiniCPM-o 4.5. The system continuously watches and listens, proactively decides when to respond, and handles interruptions on a shared causal timeline. For practitioners it offers streaming speech output, training-free long-video memory, and asynchronous tool delegation via the Realtime-Venus-Harness runtime on GitHub.
inclusionAI published Realtime-Venus on Hugging Face: a full-duplex audio-visual interaction system with two checkpoints, including a 9B Omni model adapted from MiniCPM-o 4.5. The system continuously watches and listens, proactively decides when to respond, and handles interruptions on a shared causal timeline. For practitioners it offers streaming speech output, training-free long-video memory, and asynchronous tool delegation via the Realtime-Venus-Harness runtime on GitHub.
145 engagement1 sourceredditFri, Sep 18, 2026, 03:27 PM
Releases·Model Release
NVIDIA releases NemotronLabs VoiceChat-11B full-duplex model on Hugging Face
NVIDIA has released the NemotronLabs VoiceChat-11B model on Hugging Face, featuring full-duplex voice interaction. The model enables real-time, bidirectional conversation, allowing users to speak and listen simultaneously. This release advances voice AI capabilities for developers building conversational interfaces.
NVIDIA has released the NemotronLabs VoiceChat-11B model on Hugging Face, featuring full-duplex voice interaction. The model enables real-time, bidirectional conversation, allowing users to speak and listen simultaneously. This release advances voice AI capabilities for developers building conversational interfaces.
249 engagement1 sourceredditMon, Aug 3, 2026, 10:24 PM
Releases·Model Release
Hugging Face Transformers v5.16.0 adds Qwen4-Exp with GatedResidual, QSA, PLE
Hugging Face released Transformers v5.16.0, adding support for Qwen4-Exp, a hybrid text and multimodal architecture built on Qwen3.5. The model introduces three novel components: GatedResidual (GR), Qwen Sparse Attention (QSA), and Per-Layer Embedding (PLE). These innovations aim to improve efficiency and performance in large-scale models.
Hugging Face released Transformers v5.16.0, adding support for Qwen4-Exp, a hybrid text and multimodal architecture built on Qwen3.5. The model introduces three novel components: GatedResidual (GR), Qwen Sparse Attention (QSA), and Per-Layer Embedding (PLE). These innovations aim to improve efficiency and performance in large-scale models.
0 engagement1 sourcegithubWed, Aug 26, 2026, 12:35 PM
Releases·Model Releasefeatured
Meta Superintelligence Labs releases Muse Voice Transcribe, first real-time audio perception model
Meta Superintelligence Labs launched Muse Voice Transcribe, a real-time streaming ASR model with diarization for 20+ speakers and endpointing. It is an autoregressive multimodal LLM from the Muse Spark family that processes audio in 80ms chunks and decides per chunk whether to keep listening or emit text, trained with RL on word error rate and delay rewards. It is available via the Meta Model API, Meta AI for Mac, and Muse Code.
Meta Superintelligence Labs launched Muse Voice Transcribe, a real-time streaming ASR model with diarization for 20+ speakers and endpointing. It is an autoregressive multimodal LLM from the Muse Spark family that processes audio in 80ms chunks and decides per chunk whether to keep listening or emit text, trained with RL on word error rate and delay rewards. It is available via the Meta Model API, Meta AI for Mac, and Muse Code.
0301.4k engagement3 sourcesxTue, Sep 1, 2026, 05:26 PM