Nemotron 3.5 Transcribe

Canonical version: Nemotron 3.5 Transcribe.

Nemotron 3.5 Transcribe is NVIDIA's open-weight speech recognition model, published on 28 October 2026. What sets it apart from classic ASR models like Whisper or Parakeet V3 is that it's promptable: you tell it in plain language how you want the transcript (verbatim or cleaned up, a list of terms to get right, how to format numbers), and it follows.

Under the hood, it's a speech encoder plugged into a hybrid Mamba-Transformer mixture-of-experts language model: 30B parameters in total, about 4B active per token (NVIDIA-Nemotron-3.5-Transcribe-30B-A4B). So it's a small LLM that listens, more than a pure transcription model.

What it does

• Transcription in 40 languages, with word-level timestamps
• Speaker attribution in 13 languages, up to eight speakers, in recordings up to one hour. The speaker cues reach the language model, so it tags turns while it writes, instead of in a separate diarization pass
• Translation for 31 languages, to or from English
• Audio Q&A in English: "What decisions were made in this meeting?"

One limit to keep in mind: it's "built for recorded audio, not live streaming". For dictation, you want something else.

The numbers (NVIDIA's)

• Hugging Face Open ASR leaderboard: first among open-weight models in English (5.33% average word error rate), and the lowest error in its own comparison for multilingual (4.62%) and long-form audio (10.98%)
• Artificial Analysis: 2.8% AA-WER, tied for first among open-weight models, against a model with 24B active parameters
• Multi-speaker sessions: first of 14 systems on VoiceArena's diarization benchmark (32.4% cpWER). It drops far fewer words than the others (8.32% of reference words, against a median of about 31%)
• Throughput: 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell GPU, with the upcoming NVFP4 quantized version

The fine-tuning case study is the more useful part for me. On Eka Care's Indian medical benchmark, the error rate goes from 30.8% out of the box to 23.8% with a better prompt, and to 11.8% after fine-tuning, for about 14 GPU-hours. That's a nice illustration of the order of operations: prompt first, fine-tune only when the prompt isn't enough.

Granary v2, the training data

NVIDIA also released Granary v2, an open speech dataset: 463,062 hours across 48 languages and 16 source corpora (v1 had 25 languages). Each recording comes with three versions of the transcript (normalized verbatim, with disfluencies, without), plus captions, Q&A pairs, term lists and translations. Much of it is labeled by other ASR models, then filtered (Synthetic Data at scale). Note that the release is the manifests: you download the audio yourself, under each corpus's own license.

Running it

vLLM (day one), SGLang-Omni, NeMo-Speech.cpp for local use, hosted endpoints (Baseten, DeepInfra, Simplismart), and quantized versions for DGX Spark and Jetson AGX Thor. The blog post doesn't state a license. Check the model card before using it commercially.

What the Hacker News discussion added

The thread (130+ points) was a reality check:

• Can it run without an NVIDIA GPU? Since only about 4B parameters are active, a machine with about 30 GB of RAM should run it roughly as fast as a 4B model. Someone else pointed out that porting reference Python code to your chip with a coding agent now costs $10 to $30 in tokens
• Parakeet still wins for many people. Several commenters recommend Parakeet, which "does really, really well in CPU only". A Japanese user found a Q4 quantization of Nemotron slower, heavier and less accurate than Parakeet. Others prefer Qwen ASR, or wished Mistral had open-sourced Voxtral 3
• Mixed accuracy reports. One user saw errors "every few sentences" on standard vocabulary, another saw it fail on English words in an anime theme song. On the other side, the streaming Nemotron model in the Handy dictation app "has basically solved dictation" for one user, including German mixed with English terms, down to a base M1 Mac
• Production is about the edges. Open ASR is strong now, but "things break during mulaw and turn detection" (phone-quality audio and knowing when someone stopped talking). That's where closed models still lead
• Cheap vs good. One commenter called Voxtral 3 "probably the better product", but Nemotron "cheap"

Nobody answered the most practical question in the thread: which model for thousands of hour-long interview recordings?

My take

Benchmarks say it's the best open-weight ASR model right now. The thread says: test it on your audio. Those are both true, and they don't contradict each other. Leaderboards average over clean datasets; your recordings have your accent, your jargon and your microphone (see FFASR Leaderboard for another angle on that).

What interests me is the shape of the model. Speech recognition is becoming a small LLM with ears: you prompt it, it knows who's speaking, and it can answer questions about the recording directly. For my use case (transcribing my own videos and community calls into notes), speaker labels plus a term list (Obsidian, OSK, Knowii...) in one pass would save real cleanup time. For quick dictation on my laptop, though, I'd stay with Parakeet.

References

• Hugging Face blog, "NVIDIA Nemotron 3.5 Transcribe and Granary v2" (28 October 2026): https://huggingface.co/blog/nvidia/nemotron-transcribe-granary
• Hacker News discussion: https://news.ycombinator.com/item?id=50275128

Related

• Speech-to-Text (STT)
• NVIDIA Nemotron
• Parakeet V3
• Whisper
• Voxtral Transcribe 2
• Qwen
• FFASR Leaderboard
• Artificial Analysis
• AI Mixture of Experts (MoE)
• AI Fine-Tuning
• Synthetic Data
• vLLM
• AI Open Weight Models
• HuggingFace
• Hacker News

About Sébastien

Ready to get to the next level?