NVIDIA released the Nemotron 3.5 Transcribe open-weight multilingual speech transcription model and Granary v2 data pipeline
Open source release ● Confirmed 90% confidence first seen
NVIDIA launched Nemotron 3.5 Transcribe, an open-weight multilingual speech transcription model with promptable instructions for tasks such as speaker attribution, translation, and audio Q&A. Alongside the model, NVIDIA released the Granary v2 multilingual data pipeline, and the coverage reports benchmark results on the Hugging Face Open ASR Leaderboard and claims that the model can be run locally on a single GPU.
Granary
Granary v2
NVIDIA
Nemotron 3.5 Transcribe
ElevenLabs
Hugging Face
Hugging Face Open ASR Leaderboard
NeMo Speech
Open ASR English leaderboard
The New Stack
Decision brief
- What changed
- NVIDIA released Nemotron 3.5 Transcribe, an open-weight multilingual speech transcription model, alongside the Granary v2 multilingual data pipeline. Coverage says the model supports promptable tasks such as transcription, speaker attribution, translation, and audio Q&A, and can be run locally on a single NVIDIA GPU.
- Why it matters
- This gives enterprises an open-weight alternative for multilingual speech workflows that may be easier to customize for domain-specific transcription than closed API-only options. For leaders evaluating voice AI, the combination of a reusable data pipeline, fine-tuning recipes, and local deployment claims could improve control over cost, latency, and data handling assumptions, especially where recorded-audio processing is strategic.
- Evidence
- All three cited outlets report the launch of Nemotron 3.5 Transcribe and Granary v2 and consistently describe open-weight, multilingual transcription capabilities with promptable functions. The Hugging Face blog provides the most detailed technical claims, including the 5.33% English word error rate on the Hugging Face Open ASR Leaderboard and local single-GPU deployment, while TechCrunch and The New Stack independently confirm the release and core positioning.
- What remains uncertain
- Benchmark and deployment claims in the coverage are vendor-provided or based on referenced leaderboard results, so real-world accuracy, latency, and operating cost for specific enterprise workloads remain unverified here. It is also unclear from the coverage how the model performs across all supported languages, what hardware is required in practice for production throughput, and how licensing or governance terms may affect commercial adoption.
- Monitor next
- Watch for independent enterprise benchmarks or user reports validating multilingual accuracy, production performance on single-GPU setups, and fine-tuning gains in regulated or domain-specific workloads.
Analytical support, not advice — assumptions and open questions stated above.