Nemotron 3.5 Transcribe: Czech at 5.14% WER on one GPU, with a catch for meetings

NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model that puts Czech in its top quality tier. The model card lists 13 "Leading" languages, defined as "state of the art or near state of the art out of the box", and Czech sits there next to English, German and Polish. I transcribe Czech meetings locally, so I went straight to the Czech row. It has a dash in the multispeaker column.

What the model is

The name undersells the size, and that matters the moment you pick a GPU. The card gives 34B total parameters and about 4.3B active per token, of which roughly 3.5B belong to the language backbone and about 800M to the joint audio encoder. The "30B" in the name is the backbone, taken from Nemotron 3.5 Lightning 30B-A3B, which is why TechCrunch reports 30 billion parameters while the card says 34B. Size memory from the headline and you come up 4B parameters short: in BF16, 34B is about 68 GB of weights by my arithmetic, which is why one 80 GB H100 is enough and a 48 GB card is not. Under the hood it is a Speech-Augmented Language Model, where a Transformer ASR encoder and a speaker encoder feed a hybrid Mamba-2 and Transformer mixture-of-experts backbone.

It takes recordings up to one hour with a 65,536-token context, ships in BF16 with FP8 and NVFP4 variants, and is licensed under OpenMDW-1.1, which the card describes as "ready for commercial or non-commercial use". Speculative decoding comes two ways: native multi-token prediction proposing up to 8 tokens, or an optional DFlash2 draft model. NVIDIA's reference deployment is a single H100 80 GB.

On English it posts 4.71% mean WER across the nine Open ASR Leaderboard sets, against 4.94% for Qwen3-ASR-1.7B, the next open-weight model. The multilingual leaderboard mean is 4.62% versus 6.49%, and long-form English is 10.98%.

The Czech row, read closely

Czech language identification on FLEURS is 99.45%, and speech translation from Czech to English scores 35.07 BLEU. Code-switching is marked as supported, which matters for Czech engineering calls full of English terms.

The catch is the multispeaker column. Speaker-attributed transcription, the thing that turns a recording into "who said what", is marked for English, German, French and a handful of others, but not for Czech. The model may still produce usable speaker tags on Czech audio; NVIDIA simply does not claim it, and the 15.80% cpWER it does publish comes from English meeting corpora. The card is also explicit that audio understanding and reasoning were "evaluated only in English and may not work reliably in other languages". If you want summaries straight from Czech audio, transcribe first and reason over the text with a separate model.

A leading tier on FLEURS tells you about read speech. Your meeting room is not FLEURS.

FLEURS is people reading prepared sentences. My Czech calls have crosstalk, laptop microphones and half-English jargon. Treat 5.14% as the best case, not a forecast.

Throughput, and which numbers you can reproduce

The headline efficiency claim is that one RTX PRO 6000 Blackwell Server Edition processes up to 2,238 seconds of audio per second with NVFP4 and DFlash2, returning a 30-second clip in 0.13 seconds. Read the next sentence on the card: those serving optimizations live in the NVIDIA NIM and "are not yet part of the open-source release", and the open vLLM integration is "expected to be roughly two times slower".

The numbers you can reproduce are in the vLLM section, and they are already plenty. On one H100 80 GB in BF16 with batched English ASR, NVIDIA reports roughly 410x real time with MTP2, 270x with MTP8 and 340x with DFlash2 at K=7. Take the slowest of those: an hour of audio becomes about 13 seconds of GPU time under batch load, so a week of meetings, say 40 hours, comes back in about nine minutes. Raw throughput is not the problem for a small team. Setup is where my afternoon would go: vLLM pinned to 0.32.0, the NeMo Speech main branch as a plugin, ReplaySSM flags and prefix caching off. Choose DFlash2 for speed and you also need a small patch to vLLM, and you lose word timestamps, which you need to jump from a transcript line back to the recording.

For desk hardware, the compatibility table lists the RTX 5090 with NVFP4 only and DGX Spark with all three precisions. It is an offline model; the card says internal chunking "does not make it a real-time streaming ASR service", so live captions need something else.

How I would evaluate it for Czech

I compared offline options before in meeting notes that never leave my machine, and the method carries over. Take ten real recordings, hand-correct twenty minutes of each, and compute WER on your own audio, split into single-speaker and multi-speaker segments. Run the speaker-tag prompt from the card on the Czech set and count attribution errors by hand, because no published number covers them. Budget output tokens too: the card says a dense one-hour conversation can need about 20,000, so raise max_tokens or the transcript truncates without an error. If price per hour matters more than locality, MAI-Transcribe-2 at $0.10 an hour is the hosted baseline to beat.

If your Czech audio is mostly one speaker, this is the first open model I would test against a Whisper-class pipeline. For multi-speaker Czech meetings, keep a separate diarizer in the chain until your own recordings say otherwise.