NVIDIA releases open-weight Nemotron 3.5 Transcribe: 30B MoE ASR model covering 40 languages
AccBalanced · x · 2026-10-31
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight, promptable speech transcription model built on a 30B MoE architecture with 4B active parameters.
- Transcribes 40 languages, handling recordings up to an hour
- A single checkpoint combines speaker attribution, translation, contextual prompting, and audio Q&A
- Weights available in BF16, FP8, and NVFP4
- Day-0 serving via vLLM and NeMo Speech
More from Multimodal
- Making a 10-second anime scene from character sheets with FastH3 and ComfyUI — Time-Ad-7720 · 2026-10-31
- Hunting for view-consistent image-to-video models for game character animation — Flying_Scorpion · 2026-10-31
- Fashion models looked "unexposed" in OpenAI data — then Glif launched an AI modeling agency — WillRinehart · 2026-10-31
- Cartesia launches Managed Agents; student uses it to haggle auto fare in Kannada — HeyAmit_ · 2026-10-31
- Ming-TTS inference ported to Zig/MLX hits 18x speedup on M4/M5 chips — s_busso · 2026-10-31
- 'You wouldn't think I made this with AI': another bar-raising AI video — Effective_Regular770 · 2026-10-31