October 28 catch-up
NVIDIA’s October 28 launch introduces Nemotron 3.5 Transcribe, an open-weight model for recorded audio. The card lists transcription in 40 languages, recordings up to one hour, and a 65,536-token context. Translation, speaker attribution, and audio-question support vary by task and language. It is not a live-streaming speech service. [1][2]
Active parameters are not a memory estimate
The model card describes 34 billion total parameters and about 4.3 billion active parameters, including the audio encoder. The public BF16 checkpoint listing is 67,496,013,734 bytes—about 67.5 GB. That is a file size, not a runtime RAM requirement. The card’s reference deployment uses one H100 with 80 GB of GPU memory. [1][3]
Verify the variant you can obtain
The card links FP8 and NVFP4 variants, while NVIDIA’s launch article says quantized versions are coming soon. We verified the public base-checkpoint listing but could not verify public access to those two linked variants. Check the exact artifact and supported runtime before planning a smaller machine around a promise. [1][2][3]
The checkpoint uses OpenMDW-1.1. Redistribution requires the license and applicable origin notices; other rights and permissions can still apply. Read the original terms for your intended use. [4]
Our take
This is interesting for technicians processing recordings on suitable hardware. Start with a recording you have permission to use, then check names, speaker changes, and task-specific results against the audio. An open checkpoint and an attractive active-parameter figure do not establish laptop readiness or transcription accuracy.
This briefing checks public sources. We did not run the model or reproduce vendor performance claims.