Nemotron 3.5 Transcribe Posts 4.62 Word Error Rate Across Seven Languages, NVIDIA's Own Number

NVIDIA released Nemotron 3.5 Transcribe on Oct. 28, an open-weight speech model with about 34 billion parameters in total and 4.3 billion active, covering 40 languages. The Hugging Face card says the model "is ready for commercial or non-commercial use" under the OpenMDW-1.1 licence.

The model card lists the weights on Hugging Face, NVIDIA's NeMo Speech repository on GitHub, NGC and build.nvidia.com. The name says 30B but the card lists about 34B, with the 4.3 billion active count including an audio encoder of roughly 800 million. The language backbone has about 3.5 billion active.

Word error rates, all vendor-supplied

NVIDIA reports word error rate, where lower is better. The comparison models below are the ones the card names. These are NVIDIA's own measurements.

Test	Nemotron 3.5 Transcribe	Comparison	Comparison score
English Open ASR Leaderboard, nine-dataset mean	4.71	Qwen3-ASR-1.7B	4.94
Multilingual, seven languages	4.62	Qwen3-ASR-1.7B	6.49
Long-form	10.98	Parakeet-TDT-0.6B-v3	13.37
Multispeaker (cpWER)	15.80	DiCoW v3.3 + DiariZen	16.83

1. Nemotron 3.5 Transcribe4.62 % WER
2. Qwen3-ASR-1.7B6.49 % WER

Source: NVIDIA model card for Nemotron 3.5 Transcribe, accessed 2026-11-07

On LibriSpeech the card gives 1.09 on test-clean and 2.60 on test-other. The margin on English is 0.23 points. The larger gaps are on long recordings and on seven-language speech.

One H100, one hour of audio, offline only

The recommended target is one H100 80GB. The card also lists A100, B200, GB200, DGX Spark, RTX 5090, RTX PRO 6000, H200 and L40S, depending on precision. It reports up to 2,238 seconds of audio per second on one RTX PRO 6000, using serving optimisations that are not yet open source.

Recordings can run up to one hour with a 65,536-token context. The card says the internal 30-second encoder chunking is "not a 30-second recording limit or a streaming latency claim." The model is offline. DeepInfra, a hosting provider, points streaming users to a separate 0.6B model, Nemotron 3.5 ASR Streaming Multilingual.

DeepInfra lists hosted use at $0.002 per minute, which is $0.12 per audio hour at that rate. It lists the same 34B total and 4.3B active, and repeats NVIDIA's four headline error rates.

Training data and what the card leaves open

NVIDIA says training used about 1.72 million hours of unique audio, roughly 372 million examples, or 6.53 million hours counting repeated exposure, plus 5.6 billion unique text tokens. The sources include public, third-party, vendor-licensed, web-derived and synthetic data. The card does not break that mix down by share.

Transcription and translation quality varies by language, and audio understanding was evaluated only in English. No independent evaluator's numbers appear on either page read. The final 42,605 characters of the card were not read for this report.

This is the same kind of release as the Apache 2.0 text-to-image model from Linum: open weights, a permissive licence, and vendor numbers. Mistral Large 4 drew a tighter licence for the same kind of weights.

NVIDIA has not said when the serving optimisations behind the 2,238 figure will be open-sourced.