NVIDIA Nemotron 3.5 Transcribe Review 2026: One Open Model for Transcripts, Speaker Labels and Translation
NVIDIA's new open-weight speech model transcribes 40 languages, labels speakers, translates and answers questions about recordings. Here is what is verified, what is NVIDIA's own claim, and what it really takes to run.
This NVIDIA Nemotron 3.5 Transcribe review scores it 8.17 out of 10: a strong open model on benchmarks, with real limits around live audio and Arabic dialects. NVIDIA released it on October 28, 2026 as an offline speech model that handles transcription, translation, speaker labels, word timestamps and questions about a recording from one checkpoint, steered by plain-language instructions. The weights, a companion dataset and fine-tuning recipes all shipped the same day.
The useful question is not whether it tops a leaderboard, which NVIDIA says it does among open-weight models. It is what it costs you to run, where support is thinner than the headline "40 languages" suggests, and what nobody has measured yet. We separate NVIDIA's own numbers from independent ones throughout. If self-hosting is on your list, our notes on running AI agents locally and on why a cheap GPU cloud isn't always cheap apply directly. The primary source is NVIDIA's model card on Hugging Face.
What It Can Do
Prompt-steered transcripts
Ask for verbatim text with fillers and repetitions, or a clean readable version. The instruction in the user message sets the style, and the model card says to leave the system prompt empty for transcription.
Speaker labels and word timestamps
Speaker tags travel with the transcript for up to eight speakers in 13 languages, and word-level times come back as structured output. Two catches from the docs: timestamps need a specific vLLM runner, and they are unavailable with the optional DFlash2 draft model.
Translation
31 languages translate to or from English. Several languages in the 40, including Turkish, Indonesian and Vietnamese, are transcription-only.
Context terms
Hand the model a list of names or jargon and it biases toward them. NVIDIA's own tests on contextual earnings calls show keyword recall rising from 58.16% with no context to 96.05% with local context, though those are vendor-run numbers.
Questions about audio
With reasoning switched on, you can ask what was decided in a meeting. NVIDIA evaluated this in English only and warns it may not work reliably in other languages.
A built-in fine-tuning route
The release ships Granary v2 data manifests and NeMo Speech recipes. In NVIDIA's medical case study, adapting the model took about 14 GPU-hours on one GB300 and cut held-out errors by 45.8%, so the model is meant to be customized, not just downloaded.
What It Costs to Run
The model card calls it ready for commercial or non-commercial use under OpenMDW-1.1. You pay for GPUs, power and upkeep.
NVIDIA's documented example serves BF16 on a single H100 80 GB. The compatibility table also lists FP8 and NVFP4 variants for GPUs such as the RTX 5090 and DGX Spark.
NVIDIA lists Baseten, DeepInfra and Simplismart. One third-party post reported about $0.002 a minute on DeepInfra; we could not confirm it on the provider page, so check before budgeting.
Fine-tuning uses NVIDIA's open recipes. Expect to supply your own labeled clips; NVIDIA suggests keeping at least 100 real ones for checkpoint selection.
For scale, one tester put OpenAI's Whisper API at roughly $0.006 a minute, so 1,000 hours costs about $360 at that rate. A rented GPU only beats a per-minute price if you keep it busy, which is the same trap we described in our look at serverless GPU pricing.
How the Numbers Compare
| Benchmark (lower WER is better) | Nemotron 3.5 Transcribe | Closest open-weight rival |
|---|---|---|
| English Open ASR, nine-set mean (model card) | 4.71% | Qwen3-ASR-1.7B: 4.94% |
| Multilingual Open ASR, seven languages | 4.62% | Qwen3-ASR-1.7B: 6.49% |
| Long-form Open ASR | 10.98% | Parakeet-TDT-0.6B-v3: 13.37% |
| Multi-speaker mean cpWER | 15.80 | DiCoW v3.3 + DiariZen: 16.83 |
| Artificial Analysis AA-WER (independent index) | about 2.8% | Voxtral Small (24B active): 2.77% |
Read this table with three cautions. First, NVIDIA chose the rivals and published most of the numbers. Second, the multilingual leaderboard covers only seven languages (Dutch, French, German, Hindi, Italian, Portuguese and Spanish), so the 4.62% headline says nothing about Arabic. Third, NVIDIA's comparison does not include Whisper, so we cannot say it beats Whisper.
There is also a discrepancy worth knowing about. The model card reports a 4.71% English mean, while NVIDIA's launch post cites 5.33% as its average on the live Hugging Face English leaderboard, where it ranks first among open-weight models. Check the leaderboard directly before quoting either. And the independent Artificial Analysis result is a tie with an older, larger model. Mistral's API-only Voxtral Small Transcribe 3 was reported at 2.1% on the same index by MarkTechPost, so Nemotron leads the open-weight field, not the whole market.
- We read NVIDIA's model card, including the language matrix, hardware table and benchmark appendix, and its launch post on Hugging Face.
- We cross-checked headline claims against independent coverage and the Artificial Analysis ranking that NVIDIA cites, and flagged every figure by source.
- We did not run the model ourselves. Nothing here is a claim of hands-on testing, and we found no independent hands-on test of this specific model at the time of writing.
Strengths and Weak Spots
✓ What stands out
- ✅ Open weights, commercial use allowed per the model card, with BF16, FP8 and NVFP4 variants.
- ✅ Strong published speaker-attribution results, including the lowest diarization error among 14 systems on David AI's 21-language track.
- ✅ One checkpoint for transcripts, translation, speaker labels and audio questions, steered by prompts.
- ✅ Honest documentation: the model card openly says it is offline-only and that the fastest speeds depend on an unreleased build.
- ✅ Fine-tuning is part of the package, with a worked medical example.
✗ What holds it back
- ❌ Offline only: recordings, not live streams.
- ❌ Arabic sits in the lowest "Supported" quality tier, which NVIDIA says may need fine-tuning, and no dialect-level results are published.
- ❌ "40 languages" is generous: Norwegian has no ASR checkmark in the matrix, and its FLEURS error rate is 63.76%.
- ❌ The headline speeds come from NVIDIA's optimized serving path; the open vLLM integration is expected to be about twice as slow.
- ❌ The model card notes the model can rarely fall into repetition loops and recommends a cleanup function, and it needs a pinned vLLM version.
Who Should Use It
Ideal users: Teams that transcribe meetings, calls or media in bulk, want speaker labels and timestamps from one model, and either have GPUs already or care about keeping audio on their own infrastructure. It also suits anyone who plans to fine-tune on their own jargon or accents.
Look elsewhere if: You need live captions or a voice agent (use the streaming sibling or a streaming API), you want a one-click web tool with no infrastructure, or Arabic dialect accuracy is your main requirement and you cannot spare time to test and adapt. For hands-free dictation on a laptop, tools like the one in our Willow dictation piece are closer to what you want.
Expert Editorial Opinion
The pattern in this release matters more than any single score. NVIDIA is not only handing you a transcription model. It is handing you the model, the data pipeline that shaped it, and a recipe for bending it to your own domain. That is a different pitch from "pay us per minute," and it explains why the medical fine-tuning case study sits so prominently in the launch material.
The evidence is solid but mostly NVIDIA-authored. The independent anchor is the Artificial Analysis ranking, where the model ties for first among open-weight systems at roughly one-sixth the active parameters of the model it ties with. That is the headline we would trust most, and it is worth noting it is an efficiency story as much as an accuracy story.
For Arabic speakers, the honest reading is cautious optimism. NVIDIA's numbers on read-speech benchmarks look decent: a 6.35% word error rate on FLEURS, 11.67% on CoVoST2, and near-perfect language identification. But the matrix places Arabic in the "may need fine-tuning" tier, and read speech is not a Levantine phone call. Treat it as a strong starting checkpoint you will probably want to adapt, not a finished Arabic dictation engine.
Then there is the pricing gap. The license is free; the GPU is not, and neither is the engineering time to pin versions and handle edge cases. That is the same lesson we drew in our Deepgram coverage and in the speech API consolidation experiment: the model price is rarely the bill.
Would it be worth it if NVIDIA charged per minute like a normal API? At a hosted rate near $0.002 a minute it would be competitive; at Whisper's roughly $0.006 it would clearly be. The real question is whether you can keep a GPU busy enough to beat those rates, and whether you need what a transcription API cannot give you: your own hardware, your own fine-tune, your own audio staying put. If yes, this is one of the strongest open options we have seen. If no, a hosted endpoint is the sensible way to try it first.
Final Verdict
At 8.17 it lands in the 8.0–8.9 band: strong, recommended with named caveats. The caveats are offline-only operation, Arabic in the lowest support tier, serving speeds that depend on an unreleased build, and a day-old product with mostly vendor-published evidence. For teams with GPUs and a real need for open, adaptable transcription, it is an easy shortlist entry. We did not round up for NVIDIA's name.
| Dimension | Weight | Score /10 | Why |
|---|---|---|---|
| Technical quality | 30% | 8.4 | Top open-weight results and strong speaker attribution, held back by offline-only use and thinner Arabic support. |
| Price-to-value | 25% | 8.4 | Free weights and about 4.3B active parameters, offset by GPU costs and speeds tied to an unreleased build. |
| Maturity & documentation | 20% | 7.0 | Detailed model card and tutorials, but one day old, with pinned versions and a patch for the draft model. |
| Ceiling & flexibility | 15% | 9.0 | Open weights, three precisions, self-hosting, a fine-tuning path and many tasks in one checkpoint. |
| Honesty of positioning | 10% | 8.0 | Candid about limits, but leaderboard claims are open-weight-only and two English figures (4.71% and 5.33%) circulate. |
Calculation: (8.4 × 0.30) + (8.4 × 0.25) + (7.0 × 0.20) + (9.0 × 0.15) + (8.0 × 0.10) = 2.520 + 2.100 + 1.400 + 1.350 + 0.800 = 8.17.
❓ Frequently Asked Questions
🔗 Related ToolRadar Reviews
Official sources: NVIDIA's Nemotron 3.5 Transcribe model card, the launch post with Granary v2, and the OpenMDW-1.1 license. Figures and rankings change quickly for a release this new, so confirm them there before you build.
Comments
Post a Comment