NVIDIA Launches Nemotron 3.5 Transcribe: 30B Open-Weight ASR in 40 Languages

NVIDIA Launches Nemotron 3.5 Transcribe: Free 40-Language Speech Recognition with 30 Billion Parameters

NVIDIA has officially announced Nemotron 3.5 Transcribe, a free, open-weight speech recognition model with 30 billion parameters that supports up to 40 languages, translates between 31 languages and English, distinguishes up to 8 speakers, and processes recordings up to one full hour long. The model was released on Hugging Face and announced on Wednesday, October 28, 2026, alongside the second version of the open Granary speech dataset.

Announcement Details and New Technical Features

Nemotron 3.5 Transcribe features a 30-billion-parameter architecture with partial activation (A4B) that allows it to run on just a single chip, opening the door to edge deployment via the Jetson AGX Thor robotics module, the DGX Spark chip in new laptops, and even Blackwell chips for data centers. The model belongs to NVIDIA's open-weight speech family, which also includes Parakeet at just 1.1 billion parameters, giving developers two options: an ultra-lightweight model or a heavy, high-accuracy model.

The announcement also includes the second version of the open-source Granary dataset, used to train speech models. The new tools target developers of business meeting and medical transcription applications, two of the most established AI use cases in the workplace. Developers can integrate the model with Granary and tools like NeMo Speech to customize applications for specialized terminology in each language. The one disclosed limitation: the model works with recorded audio, not live streaming.

Benchmark Comparison and Competitive Performance

NVIDIA claims that Nemotron 3.5 Transcribe outperforms competitors on the Hugging Face Open ASR Leaderboard, the reference benchmark for measuring automatic speech recognition accuracy. The launch comes during a busy period for speech releases: Mistral Voxtral 3 earlier this month, ElevenLabs Realtime for real-time speech in voice agents, and Salesforce Launa-Omni. NVIDIA's core competitive advantage is not only accuracy but also absolute free access, openness, and edge deployment capability—a combination that closed commercial models do not offer.

Industry Impact and Enterprise Adoption

For enterprises, the combination of open weights and single-chip inference changes the economics of speech AI. Organizations can deploy Nemotron 3.5 Transcribe on-premise using Jetson AGX Thor or DGX Spark hardware, reducing cloud dependency and improving data privacy for sensitive recordings such as medical consultations and internal meetings. The 40-language transcription scope and 31-language translation capability make it viable for multinational workflows, while 8-speaker diarization supports meeting transcription and multi-party interviews. The Granary dataset release further lowers the barrier to fine-tuning for domain-specific vocabulary, and integration with NeMo Speech enables customization without proprietary tooling.

Conclusion

Nemotron 3.5 Transcribe is a significant entry in the open-weight speech model market: 30 billion parameters, 40-language transcription, 31-language translation, 8-speaker diarization, hour-long audio processing, and single-chip deployment across Jetson AGX Thor, DGX Spark, and Blackwell—all free on Hugging Face. The disclosed limitation to recorded audio rather than live streaming is a constraint for real-time use cases, but for meeting, medical, and multilingual transcription workloads, NVIDIA has delivered a compelling open alternative to closed commercial APIs.

Media Source: TechCrunch AI | Official Company Statement: Original Source | Fact Verification & Analysis: AI Tools Oasis