Nemotron 3.5 Transcribe review 2026: honest scoring on accuracy, 40-language coverage, translation, diarization, availability, and who it actually fits.
Nemotron 3.5 Transcribe is one of the strongest open-weight speech-to-text models available in 2026: NVIDIA's promptable 30B-A4B model transcribes 40 languages, translates 31 to and from English, attributes up to eight speakers, and leads the open-weight tier of the Hugging Face Open ASR Leaderboard. Scored as an ASR model it is excellent and, because the weights are free, cheap to run at scale. Its limits are scope — it returns a transcript and nothing more — and that it is a model you deploy, not a finished app.
Nemotron 3.5 Transcribe is the open-weight speech model NVIDIA released on October 28, 2026, and it earned attention for two reasons: it is genuinely accurate, and it is free to download and self-host. The checkpoint (30B-A4B) is a mixture-of-experts model that turns recorded audio into text across 40 languages, translates 31 of them to or from English, and labels up to eight speakers. This review scores that model on the things that actually matter for it — how accurate the transcript is, how broad the language and translation coverage is, how good the speaker labeling is, how open and deployable it is, and what it costs to run.
I score it as what it is: an ASR model. It is not a content-creation tool, and I do not grade it as one — it writes no scripts, cuts no clips, makes no images or video, and publishes nothing. Where it competes, against other speech-to-text engines, it competes at the front of the open-weight pack, and the scores below reflect that.
Two things anchor the verdict. First, the numbers are strong: NVIDIA reports roughly a 5.33% average word error rate on English and about 4.62% across a seven-language multilingual set, leading the open-weight tier of the Hugging Face Open ASR Leaderboard, with a training recipe that draws on the open Granary v2 dataset. Second, the deliberate shape: those figures are leaderboard- and vendor-reported, the model is built for recorded audio up to about an hour rather than live streaming, and it returns text — which is the whole of its job.
Everything below reflects Nemotron 3.5 Transcribe's state as of 2026-10-28, verified against NVIDIA's launch write-up. Because it is a freshly released model, availability, tooling, and quantized builds will evolve — confirm current details and the license on the model card before you build on it.
Nemotron 3.5 Transcribe is NVIDIA's open-weight, promptable speech model, released October 28, 2026. The checkpoint, NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, is a mixture-of-experts model — roughly 30 billion parameters with about 4 billion active per token — that pairs a speech encoder with a hybrid Mamba-Transformer language model, with speaker tags produced during generation via NVIDIA's Sortformer. It transcribes 40 languages, translates 31 to and from English, attributes speakers in 13, handles recordings up to about an hour, and labels up to eight speakers. Beyond transcription it does translation, word-level timestamps, and English audio question answering. You run it as a model, not an app: the weights are free to download from Hugging Face and ModelScope, it serves on vLLM and SGLang-Omni, hosted endpoints are available through Baseten, DeepInfra, and Simplismart, and it fine-tunes with NeMo Speech tools and ms-swift. What it does is turn recorded speech into text and translations, accurately and across many languages, and that is the whole of its job.
Nemotron 3.5 Transcribe fits anyone whose bottleneck is turning recorded speech into text and who wants control over cost and deployment: teams processing a large backlog of recordings without a per-minute API bill, developers building multilingual transcription or translation into a product, and anyone who needs speaker-labeled transcripts of panels and interviews. The open weights and day-one vLLM/SGLang support make it a natural pick for self-hosting, and the 40-language transcription plus 31-language translation suit multilingual and cross-border audio. Where it fits poorly is anyone expecting a content platform: it makes transcripts and translations, not captioned video, carousels, blogs, or published posts, and there is no brand-voice layer, no clipping, and no scheduler. It is also recorded-audio first, so live captioning is the wrong job for it. If your real constraint is producing and distributing finished content rather than transcribing speech, Nemotron 3.5 Transcribe is one component of the pipeline, not the pipeline.
| Dimension | Score | Why |
|---|---|---|
| Transcription accuracy | 4.5 / 5 | NVIDIA reports ~5.33% English word error rate at a high real-time factor, leading the open-weight tier of the Hugging Face Open ASR Leaderboard — front-of-pack, though the figures are leaderboard-reported and vary with audio quality. |
| Language & translation coverage | 4.5 / 5 | Transcribes 40 languages and translates 31 to and from English — broad coverage with a real translation capability that most ASR models lack. |
| Speaker attribution (diarization) | 4.2 / 5 | Sortformer-based speaker tags for up to eight speakers, with attribution in 13 languages and strong diarization benchmarks; produced during generation rather than as a separate pass. |
| Promptability & features | 4.1 / 5 | Promptable output, word-level timestamps, and English audio question answering add flexibility beyond a plain transcript. |
| Openness / self-hostability | 4.6 / 5 | Free open weights on Hugging Face and ModelScope, serving on vLLM and SGLang-Omni from day one, and fine-tuning via NeMo Speech — excellent for teams that want to own the stack. |
| Availability / deployment | 3.8 / 5 | It is a model you deploy, not a finished app: hosted endpoints exist (Baseten, DeepInfra, Simplismart), but a quantized build for faster local inference is still listed as coming. |
| Cost & value | 4.1 / 5 | Free to download, so the running cost is your own compute — strong value for high-volume transcription, with no standalone pricing published by NVIDIA. |
| Content-workflow scope | 1.5 / 5 | Transcription and translation only — no clipping, feed captions, written content, images, video, scheduling, or publishing. Not what the model is for. |
On price, Nemotron 3.5 Transcribe is unusually easy to justify for high-volume work, because the weights are free. You download the model and the running cost becomes your own compute — a GPU you already have, a rented instance, or a per-call hosted endpoint from Baseten, DeepInfra, or Simplismart. NVIDIA did not publish standalone pricing, which is expected for an open-weight release: there is no per-minute meter unless you choose a hosted provider, so confirm the economics for your deployment path before committing.
The nuance is that "free weights" is not the same as "free to operate." Self-hosting a 30B-class model means real GPU memory and serving setup, and a hosted endpoint has its own per-call cost. For a team already running inference infrastructure, the marginal cost of transcription is genuinely low; for a solo creator who just wants clean text, a managed API may be simpler even if it meters per minute.
The honest read: as an ASR model, Nemotron 3.5 Transcribe is excellent value and, if you have the compute, close to free at the margin. What the price does not include is any of the content-production work around the transcript — cutting the clips, writing the on-brand copy, making the visuals, or publishing anything. That is not a criticism of the model; it is a reminder of scope. You are getting fast, accurate, multilingual transcription and translation, and only that.
| Use case | Fit | Why |
|---|---|---|
| Transcribing a large recording backlog | Strong | Open weights and self-hosting let you process a big archive without a per-minute API bill. |
| Multilingual and translated transcripts | Strong | 40 languages transcribed and 31 translated to and from English cover most cross-border audio needs in one model. |
| Multi-speaker panels and interviews | Strong | Speaker attribution for up to eight speakers makes long, multi-guest audio easy to navigate and quote. |
| Building transcription into a product | Strong | Free weights with day-one vLLM and SGLang-Omni support, plus hosted endpoints, make it straightforward to integrate. |
| Real-time captions / live dictation | Weak | It is built for recorded audio up to about an hour, not live streaming. |
| Turning a transcript into social posts or clips | Weak | It ends at text and translations — no clipping, no content generation, no captions burned into a feed video. |
| Producing on-brand content across platforms | Weak | No brand-voice layer, no visual formats, no scheduler — and no way to publish anything. |
To be clear where I stand: I run Kompozy, and Kompozy is not a Nemotron 3.5 Transcribe competitor. Nemotron transcribes and translates speech; Kompozy makes and publishes content. I include this note because a fair number of people evaluate an open ASR model while trying to solve a content-volume problem, and it is worth saying plainly that a transcription model will not solve that — no matter how accurate or how cheap to run, you still need something to cut the clips, write the copy, generate the visuals and video, and get it all published.
That is the honest line. If you need accurate, multilingual, self-hostable transcription — for an archive, a product you are building, or multi-speaker interviews — Nemotron 3.5 Transcribe is a genuinely strong pick and this review scores it as one. If your bottleneck is turning one recording into a week of on-brand posts across nine platforms — captioned Clipped Shorts, copy under a Persona Brief, carousels, quote cards, a blog, and a newsletter, scheduled and published from one queue — that is a content engine's job, and it is the job Kompozy is built for. Kompozy uses Whisper-based transcription internally to caption its video, so transcription is a step inside its pipeline, not the product. The clean pairing many teams land on: Nemotron 3.5 Transcribe to turn a recording (and a foreign-language one) into clean text, then Kompozy to turn that text into finished, published content.
For transcription, yes. It is an open-weight model that transcribes 40 languages, translates 31 to and from English, labels up to eight speakers, and leads the open-weight tier of the Hugging Face Open ASR Leaderboard — and because the weights are free, it is cheap to run at volume. It is not worth it as a content-creation tool, because it generates no clips, posts, images, or video and publishes nothing.
Both are open-weight ASR models you can self-host. Nemotron 3.5 Transcribe is a newer, larger mixture-of-experts model that NVIDIA reports leads the open-weight Open ASR tier, with built-in translation and speaker attribution; Whisper has a far larger tooling ecosystem and lighter hardware requirements. The right pick depends on your accuracy needs, languages, and the GPU you can run.
NVIDIA's open-weight, promptable speech model, released October 28, 2026. The 30B-A4B checkpoint turns recorded audio into text across 40 languages, translates 31 to and from English, attributes up to eight speakers, and also does word-level timestamps and English audio question answering.
The weights are free to download from Hugging Face and ModelScope, so there is no license fee to run it yourself; your cost is compute. Hosted endpoints (Baseten, DeepInfra, Simplismart) carry their own per-call pricing. Check the model card for the exact license terms.
NVIDIA reports roughly a 5.33% average word error rate on English and about 4.62% across a seven-language multilingual set, leading the open-weight tier of the Hugging Face Open ASR Leaderboard. Those are leaderboard- and vendor-reported figures; real-world accuracy varies with audio quality, so treat them as a benchmark rather than a guarantee.
No. It produces transcripts and translations and nothing more — it does not cut clips, write posts, make images or video, caption feed videos, or schedule and publish to any platform. For that you need a content engine like Kompozy, which many teams pair with a transcription model.
See NVIDIA Nemotron 3.5 Transcribe vs Kompozy comparison → · Get Started →