Nemotron 3.5 Transcribe: speaker labels from an open model, and what running it takes

Nemotron 3.5 Transcribe is NVIDIA's open-weight model that turns a recorded meeting into a speaker-labelled transcript, but it is offline-only and its model card shows its quick start on one 80 GB H100. Unless you own that hardware, try a hosted endpoint first: DeepInfra lists $0.002 per minute.

We read NVIDIA's announcement, the model card, the license and DeepInfra's page on October 29, 2026. Everything below is what those pages say that day. We did not run the model ourselves, so nothing here is a test result of ours.

Nemotron 3.5 Transcribe is an offline model, not a live captioner

NVIDIA released the checkpoint on October 28, 2026. It is a speech-augmented language model with 34B total parameters, 4.3B active, and one checkpoint covers transcription, translation, speaker labels, word timestamps and questions about the audio. NVIDIA says it transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to an hour and labels up to eight speakers.

The card is blunt about one limit: "This is an offline model." It adds that internal chunking does not make it a real-time streaming service. If you need captions while people are still talking, this is the wrong checkpoint.

NVIDIA points live use at a separate, smaller model. The Nemotron 3.5 ASR streaming card describes a 600M-parameter model with chunk sizes from 80 ms to 1,120 ms. We searched the text of that card on October 29 and found no mention of speaker labels, so do not assume the streaming model has them.

How Nemotron 3.5 Transcribe labels speakers

Speaker labels come from the prompt, not a separate pipeline. The card's example asks the model to put numbered speaker tags, spk:0 to spk:3, on the transcription, alongside instructions to include fillers and write numbers in standard form. NVIDIA says speaker cues from its Sortformer diarization encoder reach the language model with the audio, so turns are tagged while the text is generated.

For per-word times, the card says to turn on timestamp capture in the request and keep special tokens in the output. The response then carries a ctc_timestamps object with each word, its start and end in seconds, and its speaker tag. Three rules from the card are easy to miss:

• Timestamp requests accept one recording, with streaming off and a single completion.
• Check the object's error field before you trust the times.
• Keep special tokens in the output, or the speaker tags are stripped.

The card says speaker labels distinguish speakers within a recording and do not establish real-world identity. You still map "speaker 2" to a name yourself. The example prompt lists four tags, while NVIDIA says the model labels up to eight speakers, and we found no documented prompt for larger groups. Test a bigger meeting before you promise it.

Speaker labels do not cover all 40 languages. From the card's language table, they are listed for these 13:

A Dutch, Polish or Czech meeting can be transcribed, but the card's table shows no multispeaker support for those languages.

What GPU Nemotron 3.5 Transcribe needs

The card's quick start serves the full-precision checkpoint on one H100 80 GB, with vLLM 0.32.0 and the latest NeMo Speech code. The weights are a single 67.5 GB file, which is why, in our reading, a 48 GB card cannot hold the full-precision version.

The setup, in the order the card gives it:

1. Install vLLM 0.32.0 and NeMo Speech from its main branch into one Python 3.13 environment, using uv.
2. Download the checkpoint, about 67.5 GB.
3. Set the environment variables: VLLM_MAX_AUDIO_DECODE_DURATION_S to 3600, because the default is 600 seconds, and VLLM_USE_V2_MODEL_RUNNER to 0, because word timestamps need the V1 runner.
4. Start the vLLM server with the card's flags, which expose an OpenAI-compatible endpoint.
5. Send the recording with the speaker-tag prompt, temperature 0 and reasoning off, with timestamp capture on.
6. Raise the output token limit for long recordings. The card says a dense one-hour conversation can need about 20,000 output tokens.

Copy the exact commands from the card rather than from us, because the flags are long and version-pinned. Two warnings from NVIDIA belong in your plan. The card says the model can, "in rare cases", fail to stop generating and fall into a loop, and it supplies a cleanup function. NVIDIA also says the vLLM integration shipped on release day is intended for model development and expected to be roughly two times slower than its optimized NIM.

What about smaller GPUs? The card's hardware table lists the L40S 48 GB for the FP8 build and the RTX 5090 for the NVFP4 build only. But on October 29, requests for both quantized pages linked from the card returned an HTTP 401 for us, and NVIDIA's announcement says quantized versions are coming soon. We could not open either, so for now the full-precision route is the one a reader can follow.

Nemotron 3.5 Transcribe benchmarks: read the speaker numbers first

NVIDIA's headline results lead its own comparison tables, and they are NVIDIA's own tables. The card lists an English mean word error rate of 4.71% across nine public sets, ahead of 4.94% for Qwen3-ASR-1.7B. NVIDIA's announcement gives 5.33% for the English leaderboard track. The two pages do not say why the figures differ, and we did not reconcile them.

In our view, speaker attribution is where to be careful. On VoiceArena's English full-session test, NVIDIA reports the model leads at 32.4% cpWER, a metric that counts word errors after matching speaker labels. In our reading, that is roughly one reference word in three wrong, missing or attributed to the wrong person, on overlapping speech, from the best of 14 systems. On the card's four-dataset full-session average the model scores 15.80 cpWER against 16.83 for the next best published system, a modest gap. Spot-check names, speaker changes and decisions against the audio. The announcement says the same.

The card also lists a near-field prompt for dictation with background voices. It reports 4.78% WER with that prompt against 27.32% with a single-speaker prompt, on speech mixed with far-field meeting audio. NVIDIA marks that prompt experimental.

Is it better than Whisper? We cannot say. We searched the raw text of both the card and the announcement and found no mention of Whisper. Its comparison tables name Qwen3-ASR and Parakeet instead. The card says its results were captured on October 12, 2026. Run a handful of your own recordings through both before you switch.

Hosted Nemotron 3.5 Transcribe: what an hour of audio costs

DeepInfra's page lists the model at $0.00200 per minute, which is $0.12 for an hour of audio. The page shows a "Zero retention" label and requires a login to run the model. Its form includes an initial-prompt field and a chunk-level setting of segment or word. It does not say whether the hosted response returns speaker tags, so confirm that with one test file before you build on it. NVIDIA's announcement also names Baseten and Simplismart as hosts. We did not open their pages.

For scale, here are other per-minute rates from pages we read today, converted to an hour:

OpenAI labels its figures "estimated cost" for token-priced models. These rates tell you nothing about accuracy or speaker labels, so use the table only to size a bill. Forty hours of meetings a month is $4.80 at DeepInfra's rate, which makes the hardware route, in our view, hard to justify on price alone. Self-hosting makes sense when audio cannot leave your network or you plan to fine-tune.

If you want speaker labels without managing a model at all, the options in AI Audio & Voice are the better start. As of October 29, 2026, Upfynd lists 285 products there. Gladia says on its homepage that its speech-to-text API includes speaker diarization, and Sonix says on its site that its diarization identifies who said what. Those are the vendors' own claims, not tests. Rev.ai is listed too. For hosted model APIs more broadly, browse LLMs & AI Infrastructure, where we list 275 products. Our earlier post on real-time translation API prices covers the live-speech side.

Is Nemotron 3.5 Transcribe free for commercial use?

The card says the model is "ready for commercial or non-commercial use" and is governed by the OpenMDW License, version 1.1. That license grants permission "free of charge, to deal in the Model Materials without restriction". If you distribute any part of the materials, you must keep a copy of the agreement and the copyright notices. It also ends your rights if you sue someone claiming the materials infringe a patent or copyright, unless your suit answers one brought against you. And it says it "does not impose any restrictions or obligations" on outputs, so the transcripts are yours to use.

Two cautions. The card says third-party components carry their own licenses; its loop-cleanup function, for example, is Apache 2.0, not OpenMDW. And the Granary v2 training data is a separate release: NVIDIA says to download source audio under each corpus's original license.

We read these documents on October 29, 2026. Licenses and model cards change, and this is not legal advice. Read the current terms yourself, and take advice from a lawyer if real money or regulated audio, such as patient or client calls, depends on it.

Who should skip it, and the next step

Skip Nemotron 3.5 Transcribe if you need live captions, if your meetings are in a language outside the 13 with speaker labels, or if you have no GPU and no wish to run a server. In our view, a hosted app with built-in speaker labels is the better choice for most small teams, and a plain transcription API is enough when you do not need to know who said what.

Choose it if you want an open model you can fine-tune. NVIDIA's medical case study reports held-out word error rate falling from 22.44% to 12.17% after fine-tuning, and that is NVIDIA's figure on its own benchmark. Your next step: pick one recording with at least three speakers, run it through the DeepInfra endpoint, and count the wrong speaker tags by hand before you decide. If you build a transcription product on top of it, list it on Upfynd for free.