AI Expertise
Open Models / Nemotron 3.5 Transcribe
N

Nemotron 3.5 Transcribe

Visit website

Description

An open NVIDIA model for transcribing recordings: it takes audio with a text instruction and returns text. It has about 34 billion parameters in total, with about 4.3 billion active. It handles finished recordings up to an hour long. It does not transcribe live speech.

It recognizes 40 languages, including English and Russian, and recordings that switch languages. The instruction sets the style: word for word with fillers, or clean. You can add a hint with terms. The model gives word-level timestamps, identifies the language, separates speakers in some of the languages, translates speech between English and other languages and answers questions about a recording. Answering questions was tested in English only.

The full version is built for one H100 80 GB GPU, and the compressed NVFP4 version is listed for RTX 5090. It runs through vLLM and Transformers, and the authors offer an official demo on Hugging Face.

The weights are public without any acceptance step. The OpenMDW-1.1 license allows commercial use: when you redistribute, keep the license text and author notices, and your rights end if you sue over patents or copyright in the model.

Pros and cons

Pros

  • 40 languages and recordings that switch languages
  • Recordings up to an hour with word-level timestamps
  • Speaker separation and hints with terms
  • Commercial use is allowed, and the weights are public with no acceptance step
  • An official demo is available

Cons

  • Finished recordings only: no live transcription
  • Needs a server or high-end GPU
  • Speaker separation does not cover every language
  • Answering questions about a recording was tested in English only
  • The authors warn about errors, including made-up speech

Key facts

License
OpenMDW-1.1
Commercial use
Allowed
Hardware requirements
One H100 80 GB GPU; the compressed NVFP4 version is listed for RTX 5090
Developer
NVIDIA
Checked on
Nov 4, 2026

Editorial score

How strongly the AI Expertise editors recommend this open model for each of its tasks right now, on a 0–10 scale. The score combines output quality, capabilities, ease of launch, license openness and freshness.

Transcription

8.8 / 10
Output quality 9.3
Low High
Features 8.6
Basic Rich
Ease of launch 6.5
Hard Easy
License openness 10.0
Restricted Open
Freshness 10.0
Outdated Fresh

News about Nemotron 3.5 Transcribe

  • Nvidia launched open-weight Nemotron 3.5 Super VL and two lightweight Bolt models for always-on agents and edge computing. The 120B-parameter Super VL targets enterprise tool use, while Bolt runs locally on RTX Spark and DGX Spark PCs, enabling private, cloud-free AI assistants. Super VL and Bolt 74B are available; Bolt E4B comes later this month. #Nvidia #Nemotron #AIAgents #EdgeAI #OpenWeights
  • Nvidia unveiled Nemotron 3.5 Super VL, an open 120B multimodal model for nonstop enterprise AI agents. Its MoE architecture activates only 12B parameters per token, matching larger models' accuracy at one-fifth the size while parsing charts, tables, and PDFs without a separate vision model. Two smaller Bolt variants target local PCs and edge devices, including a 4GB voice-assistant version. Open weights, datasets, and training recipes are released, but everything is optimized for Nvidia hardware to drive chip sales. #Nvidia #AI #Nemotron #EnterpriseAI #OpenSource
  • Sakana AI integrates Nvidia's Nemotron models into its Fugu orchestrator, proving collective intelligence can rival single frontier models. Fugu dynamically selects specialist models for subtasks — Nemotron excels at coding, tool use, and instruction following. This modular approach reduces vendor lock-in and scales open AI through orchestration, not just bigger models. A Japanese strategy for global, resilient AI ecosystems. #SakanaAI #Nvidia #Nemotron #CollectiveIntelligence #AIOrchestration
All news about Nemotron 3.5 Transcribe
Report an error