NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, a free, open-weight speech model that transcribes audio in up to 40 languages, translates 31 of them to and from English, labels up to eight speakers, and answers questions about recorded audio — all in a single checkpoint. Alongside it came Granary v2, an open-source speech dataset whose public manifest covers 463,062 hours across 48 languages. The release targets developers building meeting, medical, and customer-service applications.
Background: Why Voice AI Is Having a Moment
Speech is one of AI's oldest commercial footholds. Transcribing a meeting, dictating a medical note, or captioning a call are tasks with obvious buyers and measurable accuracy, where getting the words right matters more than sounding clever. For years, the best transcription came from closed APIs: send audio, get text back, pay per minute.
Open-weight models change the economics. When weights are free and a model runs on a single GPU, a developer can build transcription into a product without paying a per-minute toll or sending sensitive audio — a doctor's dictation, a lawyer's deposition — to a third party's servers. That is NVIDIA's bet, and it follows a pattern: Mistral shipped Voxtral 3 earlier in October, ElevenLabs launched a real-time voice agent product, and Salesforce released Launa-Omni.
It is also a hardware story. NVIDIA dominates the chips AI runs on, and every open model that runs well on its GPUs strengthens that position. A model running on everything from a Jetson AGX Thor controller to a Blackwell data-center chip gives developers reason to stay inside NVIDIA's ecosystem from prototype to production.
What Happened
NVIDIA announced the model on October 28, 2026, via Hugging Face, with same-day coverage from TechCrunch, Silicon Report, Unite.AI, and others. The key facts:
- What it does: Transcription in 40 languages; translation of 31 languages to and from English; speaker attribution in 13 languages, labeling up to eight speakers; word-level timestamps; and audio question-answering (trained in English).
- How it is built: A 30-billion-parameter mixture-of-experts model pairing a speech encoder with a hybrid Mamba-Transformer language model, activating about 4 billion parameters per token. That sparse design lets it run on a single GPU, from edge chips to data-center hardware.
- How it is used: Users guide it with natural-language prompts — a verbatim transcript, a cleaned-up summary, or answers about the meeting — and can supply product terms or formatting conventions for specialized vocabulary in healthcare and financial services.
- Performance claims: NVIDIA says the model ranks first among open-weight models on the English track of Hugging Face's Open ASR Leaderboard, at 5.33% average word error rate, with API models excluded. Artificial Analysis reported it tied Mistral's Voxtral Small for first among open-weight models at 2.8% AA-WER, activating about 4 billion parameters versus 24 billion for the model it tied. In a company demo, fine-tuning on medical speech cut errors by a reported 45.8% after about 14 hours on one GPU. These are company-reported figures, not independently verified results.
- Speed claim: In company benchmarks, an RTX PRO 6000 Blackwell Server Edition GPU processed 100 hours of audio in under three minutes — a figure for a forthcoming quantized version, not necessarily the checkpoint available now.
- Licensing and availability: Weights are published under the OpenMDW-1.1 license, permitting free use and modification including commercial use, provided notices are kept. The model is on Hugging Face, GitHub via NVIDIA NeMo Speech, NGC, and Build.NVIDIA.com, with hosted endpoints from Baseten, DeepInfra, and Simplismart.
- Granary v2: The dataset and curation pipeline covers 463,062 hours across 48 languages, up from 25 in its predecessor. Note: the manifest describes the dataset release, not the model's full training mixture, and source audio must be downloaded separately under each corpus's original license.
There are clear limits. The model works with recorded audio, not livestreams. Speaker attribution covers only 13 of the 40 languages. Audio question-answering is trained in English. NVIDIA's materials advise spot-checking names, speaker changes, and answers against the recording.
Why It Matters for Founders and Investors
For founders, Nemotron 3.5 Transcribe lowers the cost of adding voice to a product to near zero on the model side — the remaining costs are compute, integration, and domain accuracy. That is good news for startups in healthcare documentation, legal tech, customer intelligence, and meeting productivity, where transcription has been a paid API line item. The promptable design is the interesting part: one model returning a verbatim transcript, a summary, or answers about a meeting means tiered outputs without stitching together multiple services.
The caution is dependency. An open-weight model is not an independent one: NVIDIA controls the release cadence and the ecosystem around it. Treat Nemotron as a building block, not a moat — the moat is the workflow, domain accuracy, and customer relationship built on top.
For investors, the release signals where value is migrating in voice AI. The model itself is becoming a commodity — free, open, and good enough.
Angeled.com connects founders with investors.
Silicon Report, October 28, 2026
Unite.AI, October 28, 2026
NVIDIA announcement on Hugging Face, October 28, 2026
ByteNib, October 29, 2026
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



