Nvidia debuts Nemotron 3.5 Transcribe, an open-weight model for multilingual audio transcription and translation, and Granary v2, its open-source speech dataset
Nvidia Corp. today released a leaderboard-winning open artificial intelligence model for multilingual audio transcription and translation …
Context & Ripple Effects
Nvidia has been broadening Nemotron from the 2025 family of hybrid mixture-of-experts models into a portfolio of open models for distinct workloads, including Nemotron 3 Super and the enterprise-agent-focused Nemotron 3.5 Super VL.
Transcription and translation add an audio entry point to that portfolio. Releasing both model weights and Granary v2 gives developers assets for deploying and adapting multilingual speech systems rather than only consuming a hosted API.
First-order effects
- Developers can build or tailor multilingual transcription and translation workflows with Nemotron 3.5 Transcribe and Granary v2, using Nvidia-distributed model weights and training data.
- Nvidia extends Nemotron from text and visual-agent models into speech workloads, making the family relevant to applications that begin with recorded or live audio.
Second-order effects
- Speech-AI vendors offering closed transcription APIs face a more accessible open-model baseline for multilingual deployments, particularly where customers want to control adaptation and deployment.
- Nvidia can position its broader open-model tooling more cohesively for teams assembling multimodal AI workflows after its release of NeMo Switchyard for model routing.
Third-order effects
- If Nvidia continues pairing task-specific open weights with datasets, model competition shifts from isolated benchmark models toward reusable stacks that developers can adapt for particular workflows.
- The Nemotron portfolio points to multimodal AI distribution in which speech, text and vision components are selected and composed by enterprises instead of sourced from one proprietary endpoint.
The trend: Open-model providers are expanding from general-purpose foundation models into adaptable, modality-specific building blocks for enterprise AI workflows.