Nemotron 3.5 Transcribe Replaces a Five-Model Pipeline. The Open Build Still Has Holes.

A 45-minute board call with six speakers used to require five moving parts: voice-activity detection, an ASR model, a diarizer, a forced aligner, and a language model to tidy the raw output into something a human would read. NVIDIA Nemotron 3.5 Transcribe, released on October 28, 2026 under the OpenMDW-1.1 licence, is NVIDIA's attempt to collapse that stack into a single forward pass. What makes the release worth a second read is not the collapse itself. It is the list of things the open-source serving path does not yet do.

The checkpoint is a speech-augmented language model rather than an ASR model with extras bolted on. Its 34 billion parameters are mostly a Nemotron 3.5 Lightning 30B-A3B hybrid Mamba–Transformer mixture-of-experts backbone; about 4.3 billion are active per token, including roughly 800 million of audio encoder. The repository name still reads 30B-A4B, which counts the decoder and not the assembled system. NVIDIA's own support matrix, benchmark tables and the licence all live on the model card, and that card is the most useful document in the release.

Five jobs, one checkpoint

The pitch is that a prompt chooses the transcript. NVIDIA trains five output styles into the same weights, and the caller selects one with natural-language instructions rather than a separate model or a post-processing script:

• Verbatim — disfluencies, false starts and repetitions preserved, for legal and compliance records.

• Written-form — cleaned punctuation and casing, for subtitles and readable minutes.

• Contextual — an entity list supplied in the prompt steers spelling of names, products and jargon.

• Speaker-attributed — each segment carries a speaker tag, produced inside the ASR pass.

• Rich — the kitchen-sink style that carries speaker labels, punctuation, casing and formatting together.

Beyond transcription, the same checkpoint does speech translation in both directions between English and 31 languages, spoken language identification, and audio understanding — question answering and spoken instruction following over the recording. It handles recordings up to one hour inside a 65,536-token context, using 30-second internal decode chunks that are an engineering detail rather than a recording limit. Word timestamps come from a lightweight CTC adapter after decoding.

Where the speaker labels actually come from

Most "one model does everything" claims fall apart at diarization, because speaker identity is a clustering problem and transcription is a sequence problem, and the two are usually solved separately. NVIDIA's answer is a parallel expert encoder: a frozen Nemotron 3 Diarization model produces speaker-activity representations that are injected into the speech encoder's output, so the language backbone sees who is talking while it is deciding what was said. The LLM itself is not retrained for the task.

That design is why the company can claim a single pass instead of a cascade. NVIDIA reports an average concatenated minimum-permutation word error rate of 15.80 on its multi-speaker evaluation set against 16.83 for the best specialised ASR-plus-diarizer cascade, and describes the result as the first speech LLM to beat that kind of system on speaker-attributed transcription with overlapping speech. Those are vendor numbers from a vendor-built evaluation set, and no neutral lab has reproduced them yet. Treat the direction as plausible and the margin as provisional.

The numbers, with the labels attached

NVIDIA's card reports a mean word error rate of 4.71% across the nine public English Open ASR Leaderboard test sets. The spread underneath that mean is wide and worth knowing before you quote it: LibriSpeech test-clean sits at 1.09% and SPGISpeech at 1.91%, while AMI lands at 7.81% and URGENT24 at 11.00%. The multilingual board averages 4.62% across seven languages, with German at 2.81% and Hindi at 9.15% on its own out-of-vocabulary metric.

Press coverage of the launch has also quoted 5.33% for the English board. That is a different aggregate over a different set composition, not a correction, and the two should never be placed in the same row. The card's 4.71% is the number to cite if you are citing the card.

The contextual-prompting claim is the most concrete thing in the release. On ContextASR-Bench, overall WER moves from 3.33% with no entity list to 1.71% with a fine-grained one; named-entity WER falls from 12.21% to 2.20%, and the named-entity false-negative rate from 14.99% to 1.32%. That is a real, checkable mechanism — give the model your product names and it stops guessing at them — even if the absolute scores come from NVIDIA's harness.

The independent picture is less flattering and more useful. Artificial Analysis lists Nemotron 3.5 Transcribe at 3.4% on its non-streaming AA-WER v2 table, tied around 20th to 23rd, while its own FAQ describes the model at 2.8%. Both numbers belong to the same board on the same page. The gap is a measurement-target difference — endpoint, quantisation, or evaluation slice — and it is roughly the width of the noise this index has shown across serving configurations. At the time of writing the board also lists a hosted price of $2.00 per 1,000 minutes of audio, which works out to about $0.12 per hour.

The throughput claim is the one operations teams will care about: 2,238 seconds of audio processed per second on a single RTX PRO 6000 Blackwell GPU, using the NVFP4 checkpoint and the DFlash2 draft model. NVIDIA's own caveat is that the optimised serving stack is not in the open-source release, and the shipped development integration is expected to run roughly two times slower. Halve the number before you plan capacity.

Where the open build stops

This is the part launch coverage skips. The checkpoint is capable; the integrations around it are uneven, and the model card says so plainly.

• No streaming — "This is an offline model," the card states, and adds that internal audio chunking does not make it a real-time streaming service. Every second of the recording must exist before decoding starts.

• The SGLang-Omni integration is transcription-only — it serves whole-file transcription but, per the card, does not include word timestamps, speaker-attributed output, translation or speculative decoding. The vLLM path does support timestamp capture. Which runtime you choose decides which features you get.

• Timestamps and DFlash2 are mutually exclusive — timestamps are unavailable when the speculative-decoding draft model is in use. Fast or aligned, not both, in the shipped configuration.

• Audio understanding is English-only — reasoning, spoken instruction following and audio question answering "may not work reliably in other languages," the card warns, even though transcription covers 40.

• It can fail to stop — the card recommends a cleanup guardrail because the model can occasionally loop without terminating. That is an operational requirement, not a footnote.

• One hour is a hard practical limit — performance beyond 65,536 tokens or a one-hour recording "is not guaranteed to generalize." For an all-day deposition you split the file and reconcile the speakers yourself.

Speaker attribution is also narrower than the 40-language headline: 13 of the 40 languages are marked for multispeaker output on NVIDIA's matrix, and the training mixtures cover up to eight speakers. If your meeting is in Turkish or Vietnamese, you can transcribe it well and still not get labels.

Deployment and what it costs

The weights are free for commercial and non-commercial use under OpenMDW-1.1, published in BF16, FP8, NVFP4 and GGUF, with a reference configuration on one H100 80 GB and support across Ampere, Hopper, Blackwell and Lovelace parts. There is also a NIM container for teams that would rather not assemble the serving stack, and NeMo Speech for fine-tuning — NVIDIA reports adapting the model across 17 tasks at 0.5 to 2.3 GPU-hours per task on one GPU, which is the kind of number that makes domain adaptation plausible rather than theoretical.

The economics split cleanly. Hosted, the Artificial Analysis listing of $2.00 per 1,000 minutes is about $0.12 per audio hour — cheaper than most managed transcription and roughly a third of what a typical cloud STT bill runs. Self-hosted, the marginal cost is GPU time: at the vendor's 2,238-to-1 throughput, a thousand hours of audio is under half an hour of compute on the right card, so the question stops being price and becomes whether you own or rent the hardware, and whether the audio is allowed to leave your network. Both paths are legitimate; they are not close to each other on a spreadsheet.

The half of the job the checkpoint does not do

Transcription produces text, and text is where most of the downstream cost sits: summarising the call, extracting action items, redacting personal data, translating the English output for a regional office, and checking that the speaker labels survived the handoff. Those are ordinary text-model workloads, and they do not need to run on the same box as the ASR.

That is where a router earns its place. OrcaRouter puts those downstream calls behind one key with 0% markup — the provider's list price passed through, so a vendor price change lands the same day — with automatic failover when one provider degrades and a routing DSL that can send summarisation, redaction and translation to different models inside one workflow. To be explicit: NVIDIA Nemotron 3.5 Transcribe is not in OrcaRouter's catalogue, and nothing here implies otherwise. What we route is the text layer after the transcript exists, which is the layer most teams underestimate when they budget a transcription pipeline.

If you are evaluating the model itself, the honest path is a hosted endpoint first and your own hardware second. The card's own numbers differ by more than a point between vendor and independent measurements, and the only audio that settles it is yours.

What would change the recommendation

Three things are worth watching over the next quarter. First, whether the open-source serving stack catches up to the optimised NIM path — until it does, the advertised throughput is a hosted-only number and the open release is roughly half of it. Second, whether anyone outside NVIDIA reproduces the 15.80 cpWER and the six-fold named-entity improvement; both are strong claims that currently rest on one lab's harness. Third, whether audio understanding expands beyond English, because that capability — asking questions of a recording rather than just transcribing it — is the part no competing open checkpoint offers at this size.

For now the practical verdict is narrow and unglamorous. If you need speaker-labelled, one-hour, multi-language transcription and you can accept a batch workflow, this checkpoint is the strongest open option available and the contextual-prompting mechanism is genuinely differentiating. If you need live captions, a streaming agent, or word-level alignment alongside the fastest possible decode, the release does not cover you yet, and the products that do are still the ones to call.