# Nemotron 3.5 vs Whisper: size, languages, license and WER

> NVIDIA reports 4.71% word error rate for its new open transcriber. Whisper runs on far smaller hardware. Here is how they compare on docs and published results.

URL: https://claudegpt.si/news/nemotron-3-5-vs-whisper  
Published: 2026-10-31T10:50:38.365Z  
Author: ClaudeGPT Si Newsroom  
Section: Reviews  
Topics: Nvidia, Nemotron, Whisper, OpenAI, Speech recognition  
Publisher: ClaudeGPT Si

## Key takeaways

- NVIDIA says Nemotron 3.5 Transcribe reaches 4.71% mean WER on nine English Open ASR sets, and these figures are NVIDIA's own.
- Nemotron has 34B parameters and lists an H100 80 GB for deployment, while Whisper ranges from 39M to 1.55B parameters.
- Whisper large-v3's card lists a 7.44 leaderboard WER from November 2023, which is not a like-for-like match with NVIDIA's number.

Nemotron 3.5 Transcribe and Whisper do the same job, turning speech into text, but they are built for different machines. NVIDIA's [Nemotron 3.5 Transcribe](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) is a 34B-parameter model that needs a data-center GPU and reports a lower error rate. OpenAI's [Whisper](https://github.com/openai/whisper) comes in sizes from 39M to 1.55B parameters and runs on a laptop. This Nemotron 3.5 vs Whisper comparison is based on the two projects' docs and published results, not on our own tests.

The short version: if you need the best published accuracy, many languages in one pass, speaker labels or hour-long files, Nemotron is the one to look at. If you need something small, MIT-licensed and easy to run today, Whisper is still the default. We covered the release itself in [our news story on the Nemotron launch](https://claudegpt.si/news/nvidia-nemotron-3-5-transcribe-open-speech).

## Nemotron 3.5 vs Whisper at a glance

All Nemotron figures below are NVIDIA's own, taken from its model card. Whisper figures come from OpenAI's repository and Hugging Face model cards.

| | Nemotron 3.5 Transcribe | Whisper large-v3 | Whisper turbo |
| --- | --- | --- | --- |
| Maker | NVIDIA | OpenAI | OpenAI |
| Size | 34B total, 4.3B active (mixture of experts) | 1,550M | 809M |
| Languages | 40 | Multilingual (see model card) | Multilingual, transcription only |
| Input length | Up to 1 hour, 65,536-token context | 30-second window, slides over longer files | Same |
| Extra tasks | Speaker labels, word timestamps, translation, language ID, audio Q&A | Transcription, translation, language ID | Transcription |
| Hardware | One H100 80 GB listed for deployment | About 10 GB VRAM listed for the large model | About 6 GB VRAM listed |
| Real-time | No, offline only | Not out of the box, per the model card | Same |
| License | OpenMDW-1.1 | MIT (repo), Apache-2.0 (large-v3 card) | MIT |

Whisper's VRAM figures are for the original large model in OpenAI's repo, and the speeds in that table were measured on an A100 for English. Nemotron's size is much larger, but only about 4.3B parameters are active for each token because it is a mixture of experts, which runs a subset of its network at a time.

## Which one is more accurate

NVIDIA says Nemotron 3.5 Transcribe scores a mean word error rate (WER, the share of words transcribed wrongly) of 4.71% across nine English sets from the Open ASR leaderboard. It also reports 10.98% on the leaderboard's long-form track and 4.62% on the seven-language multilingual track. NVIDIA's table calls these the best among open-weight models, and names Qwen3-ASR-1.7B and Parakeet-TDT-0.6B-v3 as the next best. Whisper is not in that table. NVIDIA says its results were captured on October 12, 2026.

Whisper large-v3's Hugging Face card lists a mean WER of 7.44 on the [Open ASR leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard), a result dated November 7, 2023. Those two numbers are not a clean head-to-head. The leaderboard has changed since 2023, NVIDIA's mean covers nine named sets including URGENT24 and Monsoon, and the two were run with different tooling. A 4.71 against a 7.44 points the same way as NVIDIA's claim, but it does not prove it. Check the live leaderboard for both models before you decide.

Whisper's card also lists known weak spots: it can write text that was never said, it can repeat itself, and accuracy drops for lower-resource languages. NVIDIA's card does not publish a matching list in the sections we read, so the absence of the same warning is not evidence the problem is gone.

## What each can do that the other can't

Nemotron takes a text instruction with the audio. NVIDIA's prompt guide shows verbatim or cleaned-up transcripts, domain vocabulary hints, speaker tags such as `<spk:0>`, translation and questions about the recording. That is closer to a speech-aware language model than to a plain transcriber.

Whisper is simpler. It transcribes, translates speech to English and identifies the language, and it has the larger tooling ecosystem around it. Its turbo variant cuts the decoder from 32 layers to 4. OpenAI's repo lists it at about eight times the speed of the large model, with a small accuracy cost. Turbo cannot translate: with the translate task it returns the original language.

## Cost, license and what you need to run it

Both are free to download. Whisper's code and weights are MIT-licensed in OpenAI's repo, while the large-v3 card on Hugging Face says Apache-2.0. Nemotron uses the OpenMDW-1.1 license, which you should read before shipping a product on it. NVIDIA's card says the model is ready for commercial use.

The hardware gap is large. NVIDIA's quick start serves the model on one H100 80 GB with vLLM. It also reports up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell GPU with its own optimized NIM serving, but says those optimizations are not yet in the open-source release and that the released integration is about two times slower. Whisper's smaller models run on a few gigabytes of VRAM.

## What we don't know yet

There is no independent, same-conditions test of the two models in the sources we read. NVIDIA's numbers are NVIDIA's claims until the Open ASR leaderboard or another third party reproduces them. We also have no published price comparison for hosted use. Watch the leaderboard for a Nemotron entry next to Whisper's.

## Frequently asked questions

### Is Nemotron 3.5 Transcribe better than Whisper?

NVIDIA reports a lower error rate and more tasks, but no independent same-conditions test of both models exists in the sources we read. Whisper is much smaller and easier to run.

### Can Nemotron 3.5 Transcribe do real-time transcription?

No. NVIDIA's model card says it is an offline model and that internal chunking does not make it a real-time streaming service.

### Is Whisper free for commercial use?

OpenAI's repo releases the code and weights under the MIT License, and the large-v3 Hugging Face card lists Apache-2.0.

## Sources

- [NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B), NVIDIA on Hugging Face
- [openai/whisper](https://github.com/openai/whisper), OpenAI on GitHub
- [openai/whisper-large-v3 model card](https://huggingface.co/openai/whisper-large-v3), OpenAI on Hugging Face
- [openai/whisper-large-v3-turbo model card](https://huggingface.co/openai/whisper-large-v3-turbo), OpenAI on Hugging Face
- [Open ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard), Hugging Face