NVIDIA Nemotron 3.5 Transcribe targets recorded audio, not live streams
NVIDIA released Nemotron 3.5 Transcribe on October 28 for recorded speech. One promptable checkpoint covers transcription, speaker attribution, translation and questions about audio. It is a model release for developers, not a new live-call transcription service.
Recorded audio, with different coverage for each task
NVIDIA’s technical announcement reports transcription in 40 languages, translation between English and 31 languages, and speaker attribution in 13. It describes support for up to eight speakers. A multilingual transcription claim therefore does not establish that every language supports every task.
The model card specifies recordings up to one hour and a 65,536-token context. Audio understanding and reasoning were evaluated only in English. Speaker labels distinguish voices within a recording, not real-world identities. Internal audio chunking does not make the model a streaming service.
Active parameters are not the whole checkpoint
The card lists 34 billion total parameters, including the audio encoder, with approximately 4.3 billion active. Its reference deployment uses one H100 with 80 GB. The smaller active count describes computation, not a promise that the full checkpoint fits in a small memory allocation.
The public file listing includes a 67.5 GB model.safetensors file alongside configuration, tokenizer and license files. That download size is not a complete estimate of runtime memory. We inspected the listing, without downloading weights or running inference.
Quantized availability is not yet clear
The launch article says quantized versions are coming soon. The model card already lists FP8 and NVFP4 variants, but our direct checks of both linked repositories returned authorization errors. We could not verify their public download availability. Do not treat a listed precision as proof of a ready-to-run laptop package.
A practical example starts with an output convention
For the same recorded meeting, NVIDIA shows requests for a speaker-labeled transcript, an English translation or a summary of decisions. Vocabulary hints can help specify names and terminology. The choice of requested output matters before any comparison of transcript quality.
An evaluation could use a short, authorized recording with a checked reference transcript. Compare names, omitted words and speaker changes separately. Then ask a question whose answer is known from the recording and verify that the response stays grounded in it. This is an editorial test plan, not a test ByteForward performed. Our Voxtral 3 migration coverage also explains why timestamp formats and live-session behavior deserve separate checks.
Model limits and server defaults are separate
The NeMo serving guide documents vLLM 0.32.0 and Python 3.11 or newer. Its default audio admission limits are 25 MB and 600 seconds. An hour-long recording can therefore be within the model’s stated capacity but exceed a server’s default limit.
| Setting | Documented default |
|---|---|
| VLLM_MAX_AUDIO_CLIP_FILESIZE_MB | 25 MB |
| VLLM_MAX_AUDIO_DECODE_DURATION_S | 600 seconds |
The guide says both limits can be raised through server environment variables. That is a configuration requirement, not evidence that arbitrary recording lengths are supported. Check the model’s limit and available resources before changing them.
The model card’s optimized throughput results use Nemotron Transcribe NIM. It says those serving optimizations are not yet in the open-source release. Those figures should not be presented as measured performance of a default local installation.
Rights and testing remain the developer’s responsibility
The OpenMDW 1.1 license grants broad use of the model materials subject to its terms. Redistribution requires retaining the license and applicable notices. Users remain responsible for third-party rights and necessary consents. Open weights do not grant permission to process someone else’s recordings.
NVIDIA links an official interactive demo. ByteForward verified its public page, but did not upload audio, test the model or independently reproduce benchmark scores.
AI-generated editorial illustration by ByteForward. NVIDIA logo reference © NVIDIA Corporation, Apache License 2.0. The logo reference was adapted into this generated illustration. No affiliation or endorsement is implied. Full image license and modification notice.
