Updated

Nemotron 3.5 Transcribe

NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech-augmented language model for instruction-guided transcription, translation, speaker attribution, and spoken-language identification across 40 languages. NVIDIA says it handles recordings up to one hour and supports word-level timestamps on request.

By No Code MBA · Facts checked

Provider
NVIDIA
Announcement date
Status at last check
Generally available

What changed

NVIDIA describes Nemotron 3.5 Transcribe as a speech-augmented language model combining a Transformer speech encoder and Nemotron 3 Diarization with the Nemotron 3.5 Lightning 30B-A3B model. NVIDIA says it supports several transcription styles, speech translation, spoken-language identification, and speaker-attributed word timestamps across 40 languages, with recordings up to one hour. NVIDIA reports approximately 4.3B active parameters and provides BF16, FP8, NVFP4, GGUF, and DFlash2 checkpoints.

Access and availability

Download the NVIDIA model checkpoints from the official Hugging Face model card. The model is designed for offline inference and supports use with NVIDIA's NeMo Speech tooling.

Official access documentation

What you can build

Evaluate it on representative recordings with domain-specific names, code-switching, and overlapping speakers. Compare verbatim and cleaned-up output before choosing a transcription workflow.

Limitations

This is an offline model rather than a hosted transcription service. NVIDIA's documented scope is up to one hour of recording and its specified 40 languages; deployment requires compatible model-serving hardware and software. Benchmark and performance claims are NVIDIA's.

Sources and release history