--- title: "NVIDIA: open full-duplex speech from a GPU company, and the licences that come with it" description: "NVIDIA released two open full-duplex speech models in 2026, PersonaPlex and NemotronLabs VoiceChat. How they were built, what NVIDIA publishes around them, from NeMo Speech and its recognisers to VideoFDB, IndicFDB and Granary, and how the same VoiceChat model can be open on Hugging Face yet evaluation-only in NVIDIA's container." article_number: "v17" slug: v17-nvidia published_at: 2026-11-10 reading_minutes: 13 tags: ["verticals", "nvidia", "full-duplex", "open-weights", "licensing"] canonical_url: https://www.fullduplex.ai/blog/v17-nvidia markdown_url: https://www.fullduplex.ai/blog/v17-nvidia/md series: "The Verticals" series_position: 17 author: "Fullduplex" site: "Fullduplex — an observatory for speech-to-speech, full-duplex & audio foundation models" license: CC BY-SA 4.0 (human) · permissive for model training with attribution --- # NVIDIA: open full-duplex speech from a GPU company, and the licences that come with it NVIDIA released two open full-duplex speech models in 2026: PersonaPlex in January, a fine-tune of Kyutai’s Moshi, and NemotronLabs VoiceChat in August, built from NVIDIA’s own recogniser and language model. Around them sit the NeMo Speech toolkit, benchmarks made with outside partners, open datasets, and a paper on handing tool calls to a separate backend model. > The company best known for GPUs now ships **open full-duplex speech models**, along with the recognisers, data and benchmarks around them. Most of it can be used commercially, but not all, and VoiceChat’s terms **depend on where you download it**: OpenMDW-1.1 on Hugging Face, evaluation-only in NVIDIA’s own container. *NVIDIA research and NeMo Speech; Two open full-duplex models, January and August 2026* ## 1. Two full-duplex models from a GPU company On January 15, 2026, NVIDIA’s Applied Deep Learning Research group (ADLR) released [PersonaPlex](https://research.nvidia.com/labs/adlr/personaplex/), a 7-billion-parameter speech model that listens and speaks at the same time, the way a phone call works. The weights went to [Hugging Face](https://huggingface.co/nvidia/personaplex-7b-v1), the code to [GitHub](https://github.com/NVIDIA/personaplex) under MIT, and the [paper](https://arxiv.org/abs/2602.06053) to arXiv the day before. PersonaPlex is a fine-tune of [Moshi](https://arxiv.org/abs/2410.00037), the open full-duplex model Kyutai released in 2024 (see [our Kyutai profile](https://www.fullduplex.ai/blog/v01-kyutai)), with two controls added: a text prompt sets the role and a voice sample sets the voice. In August, NVIDIA released a larger full-duplex model, [NemotronLabs VoiceChat 11B](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B). It uses no Moshi parts: it runs on NVIDIA’s Nemotron language model and writes tool calls on a separate output channel. The model card calls it “the first open full-duplex model to support tool calling.” The two models sit on top of a larger catalogue from the same company: the NeMo Speech toolkit, the Parakeet, Canary and Nemotron speech recognisers, a streaming diarizer, the Granary datasets, and two benchmarks for full-duplex agents built with outside partners. Few layers of an open voice agent are missing from it. > Read the licences closely. VoiceChat is under OpenMDW-1.1 on Hugging Face and under an evaluation-only licence inside NVIDIA’s NIM container, so a team that downloads the container cannot put the model into production. *Figure: NVIDIA speech releases, August 2025 to October 2026.* Dates from the model cards, arXiv listings and the NeMo Speech changelog. A month without a day means the source gives none. Audex’s card gives June 8 as its release date; its paper and Hugging Face repository date from July 6. ## 2. NeMo Speech and the recognisers Most of NVIDIA’s open speech work runs through one repository. It was created on GitHub in August 2019 as NVIDIA/NeMo and now lives at [NVIDIA-NeMo/Speech](https://github.com/NVIDIA-NeMo/Speech), under Apache 2.0, with about 18,700 stars on November 10, 2026. The project split in 2026. Its README says the repository “has pivoted to focus on audio, speech, and multimodal LLMs,” and points anyone who needs the other modalities to v2.7.3, the last release before the split. The current release, NeMo Speech 3.0, ships in the 26.07 NGC container. In August 2025 NVIDIA released [Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) and Canary 1B v2 for 25 European languages. Parakeet v3 transcribes and is under CC BY 4.0; Canary v2 also translates. On June 4, 2026, [Nemotron 3.5 ASR Streaming 0.6B](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) took NVIDIA’s English streaming recogniser to 40 language-locales in one model. Of those, 32 transcribe out of the box and the other 8 need fine-tuning first. Chunk sizes run from 80 ms to 1,120 ms and are chosen at run time. In October, [Nemotron 3.5 Transcribe](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) added an offline speech LLM for 40 languages and recordings up to an hour long. A few releases are built for voice agents in particular. [Parakeet Realtime EOU 120M](https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1) (October 2025) is a streaming recogniser with 80 to 160 ms latency that writes an `` token when the speaker finishes an utterance, so the same model that transcribes also spots the end of a turn. [Nemotron 3 Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization) (September 23, 2026) is a 100M-parameter streaming diarizer for up to eight speakers. Its card says it accepts input buffers as low as 80 ms, though it recommends 0.32 s at the low end. Both cards mark the models ready for commercial use. NVIDIA also published code that wires these parts together. The [NeMo Voice Agent](https://github.com/NVIDIA-NeMo/Speech/tree/main/examples/voice_agent) example, first released on September 5, 2025, runs NeMo speech recognition and synthesis around a Hugging Face LLM on local hardware. It has since been marked deprecated in favour of a separate repository. Other teams take the recognisers as they are. NineNineSix’s [tyger](https://huggingface.co/nineninesix/tyger), an open model that decides when a voice agent should speak, listen or backchannel, uses Nemotron 3.5 ASR Streaming 0.6B “unchanged” as its ears. *Figure: Where NVIDIA publishes in the voice stack.* The full-duplex models are in Indigo. The other cards are the layers around them, from recognition and synthesis to evaluation, data, training and serving. Only releases named in this article are listed; NVIDIA publishes more checkpoints than these. ## 3. PersonaPlex The PersonaPlex paper starts from a limit of earlier duplex models: they are “restricted to a fixed role and voice.” A customer-service agent and a general assistant need different instructions and often different voices. PersonaPlex answers with a hybrid system prompt that the model reads as one input: text that describes the role, the facts and the scenario, and an audio sample that sets the voice. Underneath, it is still Moshi. The model card names Kyutai’s Moshiko weights as the base, and the project page lists the parts: the Mimi speech encoder and decoder at 24 kHz, and temporal and depth transformers over the Helium language model. The [project page](https://research.nvidia.com/labs/adlr/personaplex/) lists three sources of training data. The real conversations are 7,303 calls, 1,217 hours, from Fisher English, two-channel telephone conversations the Linguistic Data Consortium recorded in 2003 and released as [Part 1](https://catalog.ldc.upenn.edu/LDC2004S13) (2004) and [Part 2](https://catalog.ldc.upenn.edu/LDC2005S13) (2005) under member and non-member licences. NVIDIA back-annotated each call with a prompt written by GPT-OSS-120B. The other two sources are synthetic: 39,322 assistant dialogues (410 hours) and 105,410 customer-service dialogues (1,840 hours), with transcripts from Qwen3-32B and GPT-OSS-120B voiced by Chatterbox TTS from Resemble AI. The page puts the total at “under 5,000 hours of directed data.” Real phone calls are about a third of those hours, and language models and a TTS system made the rest. That is the part we find most striking: a role-following duplex model came out of 1,217 hours of real conversation, all of it phone calls from 2003. For the released checkpoint, the model card reports a smooth turn-taking latency of 0.170 s on Full-Duplex-Bench. The model takes the turn after 95.0% of user interruptions, at 0.240 s. NVIDIA says PersonaPlex outperforms other open-source and commercial systems on these measures; the results are its own. To test roles, the team extended the benchmark to customer-service scenarios and called it ServiceDuplexBench, which the project page still says will be released “in the near future.” The code is MIT. The weights are under the [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/), which states that “Models are commercially usable” and lets you build and distribute derivative models, but ends your rights if you bypass a safety guardrail without a substantially similar replacement. The Hugging Face repository is gated, so you share contact details before downloading. The card lists English only, and the model was tested on an A100 80 GB. On November 10, 2026 the GitHub repository had about 12,100 stars and the model page 2,967 likes. ## 4. NemotronLabs VoiceChat VoiceChat reached the public in two steps. In March 2026, NeMo Speech’s changelog announced Nemotron 3 VoiceChat in early access. NVIDIA’s [early-access page](https://developer.nvidia.com/nemotron-voicechat-early-access) describes a 12B model that “unifies ASR, LLM, and TTS into a single architecture for real-time voice agents,” offered to qualified developers. Open weights followed in the summer as NemotronLabs VoiceChat 11B. NVIDIA’s own pages give different dates: the Hugging Face card says August 3, 2026, while the [NGC container listing](https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/nemotron-labs-voicechat) gives a release date of July 21 and a publication date of July 31. Every component the card names is NVIDIA’s own. A FastConformer speech encoder, taken from NVIDIA’s English streaming recogniser, feeds a Nemotron Nano v2 language model that mixes Mamba and Transformer layers. The model predicts text for an NVIDIA TTS decoder to speak, and predicts tool-calling scripts on a separate output channel. The [system report](https://arxiv.org/abs/2609.21967) (54 authors, first posted September 18, 2026) adds an RNN-T branch that transcribes the user as they speak. Training used about 550,000 hours of real and synthetic speech. The card lists Fisher again, with LibriTTS, VCTK and other speech corpora and internal Riva speaker data, plus Nemotron and UltraChat text. *Figure: PersonaPlex and NemotronLabs VoiceChat.* | | PersonaPlex | NemotronLabs VoiceChat | | ------------------------------------ | -------------------------------------------- | -------------------------------------------------- | | Released | January 15, 2026 | August 3, 2026 (model card) | | Parameters | 7B | 11B | | Built on | Kyutai Moshi (Helium LM, Mimi codec) | NVIDIA Nemotron Nano v2 LLM, FastConformer encoder | | Training audio | 1,217 h Fisher plus about 2,250 h synthetic | About 550,000 h, real and synthetic | | Control | Text prompt for role, audio prompt for voice | System prompt, plus a tool-call output channel | | Turn-taking latency | 0.170 s | 0.448 s | | Takes the turn after an interruption | 95.0% | 100% | | Weights licence | NVIDIA Open Model License | OpenMDW-1.1 | | Language | English | English | Latency and interruption figures are Full-Duplex-Bench v1 results reported by each model’s authors in its model card. They are separate runs, not a head-to-head test. PersonaPlex’s parameter count is the 7B its card gives. These numbers are NVIDIA’s own. On Full-Duplex-Bench 1.0, smooth turn-taking latency is 448 ms, and the model takes the turn after every user interruption (100%) with a GPT-4o-judged quality of 4.33 out of 5. The card ranks it “#2 amongst all open models.” The report adds that on version 1.5 it resumes after a backchannel 93% of the time, and that it scores 55.1 on VoiceBench. Tool use is the weak spot. On the tool-calling benchmark Full-Duplex-Bench v3 it scores 82.5% on tool selection, but argument accuracy is 42.2% and only 33% of tasks pass on the first try, and the authors say argument accuracy and end-to-end tool execution still need work. The model serves on vLLM, on NVIDIA GPUs from the A100 to the B200, and like PersonaPlex it is English only. The weights and the container come with different licences. On Hugging Face the weights are under OpenMDW-1.1. The NIM container on NGC, version 1.0.0, names the [NVIDIA Software and Model Evaluation License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-and-model-evaluation-license) as its governing terms. That licence limits use to “internal test and evaluation purposes, without use in production” on systems with NVIDIA GPUs, and bars disclosing evaluation results without NVIDIA’s prior written consent. To ship VoiceChat, or to publish a benchmark of it, start from the Hugging Face weights. We would check the terms on every NVIDIA artifact one by one, because the model name alone does not tell you which licence applies. The voice has a paper of its own. [VoiceChat-TTS](https://arxiv.org/abs/2608.13831) (August 13, 2026) describes itself as the speech decoder in the open VoiceChat 11B model. It reads the LLM’s text tokens as they arrive and outputs silence when there is no text. Control tokens let it be cut off mid-sentence without resetting its cache. ## 5. Handing tool calls to a backend model During 2026 several voice AI systems split talking from tool work, so the voice can keep going while reasoning and tool calls run somewhere else. LiveKit’s documentation for OpenAI’s [GPT-Live](https://docs.livekit.io/agents/models/realtime/gpt-live) describes a voice model that “handles the conversation” while “a separate backend model handles the reasoning.” On September 15, Google’s [Gemini 3.8 Live](https://blog.google/innovation-and-ai/technology/developers-tools/build-real-time-voice-applications-gemini-audio/) added tool calls that run “in the background while continuing to stream audio responses to the user.” Ant Group and Tsinghua University released [Realtime-Venus](https://huggingface.co/inclusionAI/Realtime-Venus), two open 9B models under Apache 2.0 that emit in-stream `` requests to a separate harness. NVIDIA wrote the pattern up as an architecture. Its [frontend-backend paper](https://arxiv.org/abs/2609.19334), with 13 authors and first posted on September 16, trains a duplex speech-to-text frontend to emit a delegation token and pass its streaming transcript to a text LLM, which makes the tool calls. Results come back to the frontend through a light prefill-and-repeat step and reach the user through streaming TTS. The frontend barely changes, and the authors report that turn-taking and interruption handling are largely kept. In a single-turn test, the system recalls 92 to 97% of needed tool calls and correctly rejects irrelevant ones 81.2% of the time. With a larger backend, Qwen3-235B-A22B, it is competitive on Full-Duplex-Bench v3 and outperforms GPT-realtime-mini and Qwen3-Omni-30B-A3B-Instruct on EVA-Bench. The authors conclude that “backend delegation is an effective and modular approach.” *Figure: Two ways to give a full-duplex model tools.* Left: VoiceChat writes tool calls as a second output stream of the same model. Right: the frontend-backend design keeps the duplex model small and sends tool work to a text LLM behind it. Both are NVIDIA designs, published about six weeks apart. Simplified from the model card and the two papers. So NVIDIA has published both designs. VoiceChat puts tools inside one model, where the voice and the tool call share weights and a single serving stack. The frontend-backend paper keeps the model that decides when to speak small and lets a team swap the model that decides what to do. Even VoiceChat keeps its voice as a separate module, and the VoiceChat-TTS paper presents that as responsive, always-on speech that keeps modularity and speech quality. NVIDIA has not said that one design replaces the other. The papers do hint at the trade. VoiceChat’s authors name argument accuracy as its weak point, while the frontend-backend paper gets its Full-Duplex-Bench v3 result by plugging in a larger backend, an option the one-model design does not have. If we needed dependable tool calls today, we would start with the delegated design and accept a second model to run. The one-model route is simpler to serve, but on its authors’ own numbers it is not yet reliable with tool arguments. ## 6. Benchmarks, data and licences **Benchmarks built with partners** NVIDIA’s two full-duplex benchmarks were both built with outside organisations. [VideoFDB](https://research.nvidia.com/labs/amri/projects/video-fdb/) (arXiv, [May 28, 2026](https://arxiv.org/abs/2605.30256)) comes from NVIDIA and David AI and calls itself “the first benchmark to evaluate full-duplex audio-visual-to-audio-visual (AV2AV) conversational agents.” It uses 237 clips from real two-person video calls covering 11 nonverbal dynamics, such as backchannels and gaze aversion, and a language-model judge scores each on 0 to 5 rubrics for perception and generation. The authors found two failure modes in current agents, “captioning collapse” and “visual-stream ignorance,” and conclude that the architecture of cascaded speech-to-avatar systems rules out full-duplex nonverbal cues. Its [leaderboard](https://research.nvidia.com/labs/amri/projects/video-fdb/leaderboard.html) takes uploaded agent outputs through an evaluation Space. [IndicFDB](https://arxiv.org/abs/2609.31967) (September 25, 2026), from NVIDIA and VoiceArena, extends Full-Duplex-Bench to ten languages: Indian English and nine Indian languages from Hindi and Bengali to Tamil and Malayalam. Its 12,350 samples were mined from about 50,000 hours of channel-separated conversation. The seven agents tested include PersonaPlex, Moshi and four commercial APIs. Those APIs behave consistently across languages, the paper finds, but each is either fast or tolerant of pauses, not both. The paper describes the samples and code as intended for release. The same researchers recur across these releases. Rajarshi Roy is an author of PersonaPlex, VideoFDB, IndicFDB and VoiceChat-TTS. Jaehyeon Kim is on PersonaPlex, VoiceChat-TTS and [Audex](https://arxiv.org/abs/2607.05196), and Bryan Catanzaro on PersonaPlex, IndicFDB and Audex. **Datasets** NVIDIA’s open speech datasets are large, but they serve recognition and translation. [Granary](https://huggingface.co/datasets/nvidia/Granary) ([May 2025](https://arxiv.org/abs/2505.13404)) covers 25 European languages, with a headline figure of about 1 million hours. Its speech recognition part is about 643,000 hours: 192,000 ship with audio, and 451,000 are new transcripts for corpora you download from their original sources. The rest is speech translation into English. The pipeline is open in NeMo’s Speech Data Processor. [Granary-v2](https://huggingface.co/datasets/nvidia/Granary-v2) ([October 28, 2026](https://research.nvidia.com/labs/nemotron/Granary-v2/)) publishes labels for 463,062 hours of source speech in 48 languages from 16 corpora, about 165,300 hours of it flagged high quality, under OpenMDW-1.1 and CC BY-SA 4.0. The project page puts the full processed collection at about 843,600 hours. It is part of the training data for Nemotron 3.5 Transcribe. By the card’s own table, 380,711 of Granary-v2’s hours come from [VoxPopuli](https://github.com/facebookresearch/voxpopuli), recordings of the European Parliament, and most of the rest from audiobooks. In both, one person speaks at a time. For two-person conversation, NVIDIA’s full-duplex models went outside the company: both list Fisher’s 2003 phone calls among their training data, alongside synthetic speech. This is the gap we would most like to see closed. NVIDIA publishes speech labels by the hundreds of thousands of hours, yet the real conversations its own duplex models learn from are more than twenty years old and come from someone else. **Licences** On [May 28, 2026](https://www.linuxfoundation.org/press/linux-foundation-releases-openmdw-1.1-nvidia-adopts-openmdw-for-cosmos-isaac-gr00t-ising-and-nemotron-ai-model-families) the Linux Foundation released OpenMDW-1.1, one licence that covers a model’s weights, code, documentation and data, and announced that NVIDIA would adopt it for future releases of its Cosmos, Isaac GR00T, Ising and Nemotron families. “Open innovation is foundational to AI progress,” said Kari Briski, NVIDIA’s vice president of generative AI. Several speech releases since then carry it: Nemotron 3.5 ASR Streaming in June, the VoiceChat weights in August, Nemotron 3 Diarization in September, and Nemotron 3.5 Transcribe and Granary-v2 in October. PersonaPlex stays under the NVIDIA Open Model License. [Nemotron-Labs-Audex 30B-A3B](https://huggingface.co/nvidia/Nemotron-Labs-Audex-30B-A3B), published with its paper in July 2026, handles audio understanding, recognition, translation, TTS and speech-to-speech in one network. It uses the NVIDIA OneWay Noncommercial License, which allows use “non-commercially,” defined as “for research or evaluation purposes only,” while NVIDIA and its affiliates may use the work and its derivatives commercially. The VoiceChat NIM container is evaluation-only. *Figure: Licences across NVIDIA's speech releases.* | Licence | Commercial use | Speech releases in this article | | ------------------------------------ | ------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------- | | Creative Commons BY and BY-SA | Yes, with attribution; BY-SA adds share-alike | Parakeet TDT 0.6B v3 (BY 4.0); Granary (BY 4.0 and BY 3.0 parts); Granary-v2 (BY-SA 4.0, with OpenMDW) | | OpenMDW-1.1 | Yes: use without restriction if the licence is kept; rights end if you sue claiming the model infringes | VoiceChat 11B weights; Nemotron 3.5 ASR Streaming; Nemotron 3 Diarization; Nemotron 3.5 Transcribe; Granary-v2 | | NVIDIA Open Model License | Yes, on NVIDIA's terms; rights end if a guardrail is bypassed without a substitute | PersonaPlex weights; Parakeet Realtime EOU 120M; Nemotron Speech Streaming En 0.6B; MagpieTTS multilingual 357M | | NVIDIA OneWay Noncommercial | No: research or evaluation only, while NVIDIA keeps commercial use | Nemotron-Labs-Audex 30B-A3B | | NVIDIA Software and Model Evaluation | No: internal evaluation on NVIDIA GPUs; results need NVIDIA's consent to publish | VoiceChat NIM container on NGC | Grass: commercial use allowed on standard open terms. Amber: allowed, on NVIDIA’s own conditions. Tomato: research, evaluation or internal testing only. Code is licensed separately: PersonaPlex under MIT, NeMo Speech under Apache 2.0. The NGC page for the VoiceChat container also says its components are ready for commercial or non-commercial use; the governing terms it names are the evaluation licence. Read each licence before use; this table is a summary. **Open items** Every benchmark number for the two models in this article comes from NVIDIA’s own cards and papers, so how they hold up in other people’s tests is still unproven. Two announced pieces are worth checking for. The PersonaPlex page, dated January 15, describes ServiceDuplexBench as coming “in the near future,” and the IndicFDB paper describes its samples and code as intended for release. Language is the other gap. Both full-duplex models are English only, while the streaming recogniser covers 40 language-locales and IndicFDB already tests agents in ten languages spoken in India. ## Sources - PersonaPlex: [project page](https://research.nvidia.com/labs/adlr/personaplex/), [paper](https://arxiv.org/abs/2602.06053), [model card](https://huggingface.co/nvidia/personaplex-7b-v1), [code](https://github.com/NVIDIA/personaplex) - NemotronLabs VoiceChat: [model card](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B), [system report](https://arxiv.org/abs/2609.21967), [NGC container](https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/nemotron-labs-voicechat), [early access page](https://developer.nvidia.com/nemotron-voicechat-early-access), [VoiceChat-TTS](https://arxiv.org/abs/2608.13831) - Delegation: [NVIDIA frontend-backend paper](https://arxiv.org/abs/2609.19334), [LiveKit GPT-Live docs](https://docs.livekit.io/agents/models/realtime/gpt-live), [Google on Gemini 3.8 Live](https://blog.google/innovation-and-ai/technology/developers-tools/build-real-time-voice-applications-gemini-audio/), [Realtime-Venus](https://huggingface.co/inclusionAI/Realtime-Venus) - NeMo Speech and recognition: [NeMo Speech](https://github.com/NVIDIA-NeMo/Speech), [NeMo Voice Agent](https://github.com/NVIDIA-NeMo/Speech/tree/main/examples/voice_agent), [Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3), [Nemotron 3.5 ASR Streaming](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b), [Parakeet Realtime EOU](https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1), [Nemotron 3 Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization), [Nemotron 3.5 Transcribe](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B), [Nemotron Speech Streaming En 0.6B](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b), [MagpieTTS multilingual 357M](https://huggingface.co/nvidia/magpie_tts_multilingual_357m), [tyger](https://huggingface.co/nineninesix/tyger) - Audio understanding: [Nemotron 3 Nano Omni](https://arxiv.org/abs/2604.24954), [Audex model card](https://huggingface.co/nvidia/Nemotron-Labs-Audex-30B-A3B), [Audex paper](https://arxiv.org/abs/2607.05196) - Benchmarks: [VideoFDB](https://research.nvidia.com/labs/amri/projects/video-fdb/), [VideoFDB paper](https://arxiv.org/abs/2605.30256), [IndicFDB](https://arxiv.org/abs/2609.31967) - Data: [Granary](https://huggingface.co/datasets/nvidia/Granary), [Granary paper](https://arxiv.org/abs/2505.13404), [Granary-v2](https://huggingface.co/datasets/nvidia/Granary-v2), [Granary-v2 project page](https://research.nvidia.com/labs/nemotron/Granary-v2/), [Fisher English Part 1](https://catalog.ldc.upenn.edu/LDC2004S13), [Part 2](https://catalog.ldc.upenn.edu/LDC2005S13) - Licences: [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/), [NVIDIA Software and Model Evaluation License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-and-model-evaluation-license), [OpenMDW-1.1 text](https://openmdw.ai/license/1-1/), [Linux Foundation on OpenMDW-1.1](https://www.linuxfoundation.org/press/linux-foundation-releases-openmdw-1.1-nvidia-adopts-openmdw-for-cosmos-isaac-gr00t-ising-and-nemotron-ai-model-families) --- _Originally published at [https://www.fullduplex.ai/blog/v17-nvidia](https://www.fullduplex.ai/blog/v17-nvidia)._ _Part of **The Verticals** · v17 / 26 · from Fullduplex._ _Full index: https://www.fullduplex.ai/blog · Markdown of every article: https://www.fullduplex.ai/llms-full.txt._