nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
just landedmodel
Nemotron 3.5 Transcribe is an open-weight, offline speech and audio model for transcription in 40 languages, with support for recordings up to one hour and a 65,536-token context. Mixture-of-experts, 34B total, 4B active (128 experts, 6 per token). Context: 256K tokens (config). License: openmdw-1.1.
- Org
- nvidia
- Released
- 2026-10-07
- Status
- just landed
- License
- openmdw-1.1 (permissive)
- Parameters
- 33.7B total, 4.0B active
- Architecture
- MoE · 128 experts, 6/token · Nemotron3_5TranscribeForConditionalGeneration
- Context
- 262,144 tokens
- Task
- automatic-speech-recognition · audio
- Signal
- 175 HF likes · 338 downloads
Runs on
mixed: real files + estimate; need = weights + FP16 KV cache for 8K tokens + 1GB. A sizing calculation, not tested.
- Q4 ~24.1 GB (estimated size): 32GB GPU (5090), Mac 36GB, Mac 48GB, CPU only, 32GB RAM (slow)
- Q8 ~38.9 GB: 48GB GPU (RTX 6000 Ada / A6000), Mac 64GB (M1 Max), CPU only, 64GB RAM (slow)
- FP16 ~71.3 GB (estimated size): 96GB GPU (RTX Pro 6000 Blackwell), 1× DGX Spark (128GB), 2× DGX Spark (256GB), Strix Halo 128GB unified (~96GB usable), Mac 96GB, Mac 128GB, Mac 192GB, Mac 256GB, Mac 512GB, CPU only, 128GB RAM (slow), CPU only, 256GB RAM (slow)
Links
First published here 2026-10-07 · updated 2026-10-28 · markdown · JSON