![Nemotron 3.5 Transcribe Fits 17 GB, 5.33% WER [2026]](https://shattered.io/wp-content/uploads/2026/10/nemotron-3-5-transcribe-17gb-nvfp4-5-33-wer-self-hosting-2026-gen.webp)
Nvidia released Nemotron 3.5 Transcribe on October 28, 2026, an open-weight speech model that transcribes 40 languages, translates 31 of them to or from English, and labels speakers in 13. TechCrunch, The Decoder and Unite.AI all reported the launch within hours. The headline spec is not the language count. It is the memory footprint. According to Tech Times, the NVFP4 build needs roughly 17 GB of VRAM, which puts a 40-language speech stack on a single mid-range card.
We already covered the basic launch details in our earlier Nemotron 3.5 Transcribe report. This analysis goes further. It works through the self-hosting math, separates the benchmark numbers that agree from the ones that conflict across outlets, and places the release inside a very crowded month for voice AI. Where reports disagree, we say so and use the more conservative figure.
One caveat shapes everything below. Nemotron 3.5 Transcribe works on recorded audio, not live streams. That single limit decides which products can adopt it today and which cannot. We return to it in detail later.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What Nvidia Shipped on October 28 #
The release has two halves. The first is the model itself, a speech LLM that follows text instructions. You can ask it to transcribe, to translate, to attribute speakers, or to answer a question about the audio. The second half is Granary v2, an open speech dataset and curation pipeline that Nvidia published alongside the model. Unite.AI reports that the public manifest covers 463,062 hours across 48 languages.
Distribution is broad. Unite.AI lists Hugging Face, GitHub, NGC and Build.Nvidia.com as launch locations, and Tech Times adds ModelScope. Hosted inference is available from Baseten, DeepInfra and Simplismart, per Unite.AI. For self-hosting, both outlets point to vLLM and the NeMo Speech tooling, with Unite.AI naming vLLM 0.32.0 plus a NeMo Speech plugin.
The model ships in three precisions: BF16, FP8 and NVFP4. That matters because each one maps to a different class of hardware. We break that down in the VRAM section.
The AI Insider, MarkTechPost, Unite.AI and Tech Times covered the release within a day of each other. TechCrunch framed it as part of a busy stretch for voice AI. That framing is accurate, and it explains why Nvidia chose to ship a model that is free to download.
Nemotron 3.5 Transcribe Specs at a Glance #
The table below collects the figures that two or more outlets agree on, plus the ones where a single outlet supplies the detail. Where a number conflicts, the right-hand column flags it.
| Spec | Reported value | Source and note |
|---|---|---|
| Release date | October 28, 2026 | TechCrunch, MarkTechPost, Tech Times |
| Languages transcribed | 40 | All outlets agree |
| Translation | 31 languages to or from English | The Decoder, Unite.AI, TechCrunch |
| Speaker attribution | 13 languages, up to 8 speakers | The Decoder, Unite.AI |
| Max recording length | One hour | TechCrunch, Unite.AI |
| Parameters | About 30B to 34B total, 4B to 4.3B active | Outlets differ, see architecture section |
| Context window | 65,536 tokens | Unite.AI |
| License | OpenMDW-1.1 | Tech Times, Unite.AI |
| Precisions | BF16, FP8, NVFP4 | Tech Times |
| Streaming | Not supported | TechCrunch, Tech Times |
Two details stand out. The one-hour ceiling and the eight-speaker cap describe a meeting recorder, not a call-center agent. And the 65,536-token context window tells you why the hour limit exists. Audio gets converted into tokens, and an hour of speech has to fit.
Architecture: Why 4.3B Active Parameters Matter #
MarkTechPost describes Nemotron 3.5 Transcribe as a mixture-of-experts model with about 30 billion total parameters and roughly 4 billion active per step. Unite.AI and Tech Times give 34 billion total and about 4.3 billion active per token. The Decoder lands on the rounder 30 billion and 4 billion figures.
Which Parameter Count to Trust
The memory numbers favor the larger count. Tech Times puts BF16 at approximately 68 GB, and 34 billion parameters at two bytes each comes to 68 GB. That arithmetic matches the 34B figure. The 30B number looks like a rounded version of the same checkpoint. We use 34B total and 4.3B active when we need precision, and we flag the discrepancy when we do.
The active count is the number that drives speed and cost. In a mixture-of-experts design, each token only passes through a subset of the weights. The full set must sit in memory, but compute per token scales with the active slice. That is why Tech Times says the model matches rivals using one-sixth the active parameters. It compares Nemotron’s 4.3B against Mistral’s Voxtral Small, which Tech Times puts at roughly 24 billion active.
Our own arithmetic agrees with that framing. Four billion divided into twenty-four billion is one-sixth. With 4.3B the ratio sits closer to 18 percent, so one-sixth is a fair round number rather than an exact one.
Benchmarks: 5.33% WER and the Fine Print #
Word error rate, or WER, counts the share of words a model gets wrong against a human reference. Lower is better. Nemotron 3.5 Transcribe posts strong numbers, but they come from several different test sets, and not every outlet reports the same value.
| Benchmark | Reported result | Source and caveat |
|---|---|---|
| Hugging Face English Open ASR Leaderboard | 5.33% average WER, first among open-weight models | The Decoder, Unite.AI. Tech Times reports 4.71% |
| Multilingual mean (seven languages) | 4.62% WER | Unite.AI, citing Nvidia |
| Long-form audio | 10.98% WER among open-weight models | Unite.AI |
| Artificial Analysis AA-WER | 2.81% (Voxtral Small: 2.77%) | The Decoder, Tech Times |
| VoiceArena speaker diarization | 32.4% cpWER, lowest of 14 systems | The Decoder |
| David AI multilingual (DAI-ASR) | 18.26% error rate | The Decoder, Tech Times |
The 5.33 percent leaderboard figure appears in two outlets, so we treat it as the working number. Tech Times reports 4.71 percent for the same leaderboard, which may reflect a different snapshot or a different averaging. We cannot resolve that from the published coverage. Readers who plan to cite the figure should check the live leaderboard.
The Voxtral Small Comparison
The cleanest head-to-head is the Artificial Analysis index. Nemotron scores 2.81 percent and Voxtral Small scores 2.77 percent, a gap of 0.04 points. Mistral’s model is slightly ahead on that test. Nemotron gets there with far fewer active parameters, so the story is efficiency, not a raw accuracy win. Anyone who reads this release as Nvidia beating Mistral on accuracy is reading the wrong table.
Diarization is where Nvidia claims a clear lead. The Decoder reports 32.4 percent cpWER on VoiceArena, the lowest among 14 systems tested. cpWER penalizes errors in who said what as well as what was said, so it suits meeting transcripts. A 32.4 percent score still means roughly one word in three is wrong or misattributed on that hard test, which is a reminder that overlapping speech remains difficult for every system.
VRAM Math: What It Takes to Run It Yourself #
Tech Times publishes three memory figures, one per precision. They line up with the usual rule of thumb. Halving the bits per weight roughly halves the memory.
| Precision | Approx. VRAM (Tech Times) | Practical read |
|---|---|---|
| BF16 | About 68 GB | Needs a workstation or datacenter card with 80 GB-class memory |
| FP8 | About 34 GB | Fits a 48 GB-class professional card |
| NVFP4 | About 17 GB | Fits a 24 GB-class card with room for audio buffers |
The right-hand column is our inference, not a vendor claim. These figures cover the weights and basic overhead. Real deployments add memory for the key-value cache, and long recordings push that up. A card with only a few gigabytes of headroom over the weight footprint will struggle with hour-long files and batching.
The Decoder says the model runs on single chips from Nvidia’s range, and TechCrunch names Jetson AGX Thor, DGX Spark and Blackwell datacenter parts. If you want a feel for the class of desktop hardware this targets, our RTX Spark launch review roundup covers a 128 GB unified-memory machine that clears even the BF16 footprint. NVFP4 is a Blackwell-oriented format, so the smallest build is also the one that ties you most closely to Nvidia’s newest silicon. Teams on older GPUs should plan around FP8 or BF16.
Throughput: 100 Hours in Under Three Minutes #
The Decoder reports that the model processes 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell. Do the math. One hundred hours is 6,000 minutes. Dividing by three gives a floor of 2,000 times real time. That is our arithmetic on the reported figure, and the true speed is at least that.
For a back catalog, that changes the economics. A media archive with 10,000 hours of recordings would take roughly 5 hours of GPU time at the reported pace, assuming the speed holds across a mixed set. Real batches vary, so treat that as a planning estimate rather than a promise. The vendor did not publish a cost per hour, and we will not invent one. Throughput on recorded audio is exactly where a model without streaming support can still win. Batch transcription, compliance review, subtitle generation and meeting archives all reward speed over latency. None of them care whether the first word arrives in 200 milliseconds.
Speaker Labels, Translation and Instruction Control #
Most open ASR models do one job. They turn audio into text and stop. Nemotron 3.5 Transcribe takes instructions in plain text, according to The Decoder, which lets one checkpoint cover several tasks. You can request a transcript, a translation, speaker attribution, or an answer about the audio.
The Decoder also lists technical term hinting and handling of overlapping speech. Hinting lets you pass in a vocabulary list, such as drug names or product codes, so the model leans toward them. Tech Times reports a domain-specific baseline of 30.8 percent WER, 20.8 percent with prompting, and 11.8 percent after fine-tuning. Those three numbers show how much of the gain comes from cheap prompt changes and how much needs training.
Unite.AI reports a different medical case study, with WER falling from 22.44 percent to 12.17 percent on held-out clips. The two outlets may describe different experiments. Either way, the direction is consistent. Domain fine-tuning cut error rates by about half or more in the cases reported.
Audio question answering exists but is limited. Tech Times says it works primarily in English and calls non-English question answering supplementary rather than a primary capability. That honesty from the model card is useful. It tells builders not to promise multilingual audio question answering to customers.
Granary v2: The Dataset Behind the Model #
The dataset deserves its own look because it may outlast the model. Unite.AI reports Granary v2 spans 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in the first version. Nvidia built it with NeMo Curator, and it generates several transcript variants plus translations with quality scores.
The filtering numbers show how aggressive the cleanup was. Unite.AI says the pipeline cut VoxPopuli from 354,000 hours to 111,000 hours across 22 languages. That is a reduction of about 69 percent. Nvidia is betting that less audio with cleaner labels beats more audio with noisy ones.
There is a licensing wrinkle. Unite.AI notes that source audio is downloaded separately under each corpus’s original license. Nvidia ships the manifest and the tooling, not the audio itself. Anyone building a commercial product on retrained weights should read each source license before shipping.
For smaller language communities, the open pipeline may matter more than the checkpoint. Public-sector projects such as [India’s CIIL Anuvani speech effort for 22 languages](https://shattered.io/ciil-anuvani-open-source-speech-ai-22-indian-languages-2026/) show the demand for open speech data outside the big European and East Asian markets. A reusable curation recipe lowers the cost of those projects.
## OpenMDW-1.1: What Open-Weight Means Here
Tech Times and Unite.AI both report the OpenMDW-1.1 license. Unite.AI says it permits commercial and non-commercial use. We have not seen the full license text in the coverage, so we cannot describe its edge cases, such as attribution duties or restrictions on derivative works. Readers who plan a commercial rollout should read the text on the model page before committing.
The term open-weight is narrower than open source. You get the trained parameters and can run, fine-tune and host them. You do not necessarily get the full training recipe or the right to relicense the model. Nvidia does publish Granary v2 tooling, which pushes this release closer to the open end than many weight-only drops.
Our comparison of six open-weight model families covers how license terms differ across the field. The practical lesson there applies here too. A permissive headline hides the clauses that matter most when a company scales past a few thousand users.
A Crowded Month: Voxtral 3, ElevenLabs and Launa-Omni #
TechCrunch placed the release in what it called a busy period for voice AI. It names three rivals. Mistral released Voxtral 3 earlier in October 2026, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released a voice model called Launa-Omni. We have no benchmark data comparing those models to Nemotron beyond the Voxtral Small figures above, so we do not rank them.
The pattern is still clear. Voice is splitting into two product lines. One line serves live conversation, where latency is everything. ElevenLabs Realtime sits there. The other line serves recorded audio, where accuracy, speaker labels and cost per hour decide the winner. Nemotron 3.5 Transcribe sits squarely in that second group.
Mistral is Nvidia’s most direct open-weight competitor in this segment. Its activity this month extends beyond speech, as we noted in our look at Mistral Large 4 going open. Two European and American labs are now trading open releases in close succession. That rhythm favors developers, since each release pressures the other side on price and license terms.
Market Impact: Who Gains and Who Loses #
The clearest winners are teams that already pay per minute for transcription. A free checkpoint that runs on one card sets a ceiling on what managed services can charge for batch work. Hosting providers named at launch, Baseten, DeepInfra and Simplismart, gain a new product to sell without building the model. They will compete on price and uptime rather than on accuracy.
Nvidia gains in a different way. The model is free, but the best builds are tuned for its hardware, and NVFP4 favors Blackwell. Releasing strong open models creates demand for the chips that run them best. This is the same logic that has driven its open model program throughout 2026, and it connects to the capital questions we covered in Nvidia’s reported credit backing for a $45 billion Nscale deal.
Pure-play speech vendors face the hardest pressure. Companies that sell accuracy alone, with no workflow layer, now compete against a free model that scores near the top of public leaderboards. Those with strong integrations, compliance tooling, or live-call products have more room. For wider coverage of where the money is moving, see our AI and machine learning section.
Historical Context: From Whisper to Speech LLMs #
OpenAI’s open-source Whisper set the baseline for free transcription and became the default choice for a generation of apps. Its appeal was simple. One download covered many languages with no per-minute fee. The weakness was equally familiar to developers. Whisper did not label speakers on its own, and teams stitched on separate diarization tools.
Nemotron 3.5 Transcribe reflects the next step in that arc. Speech models now behave like language models. They take instructions, accept hints and handle several tasks from one set of weights. The Hugging Face Open ASR Leaderboard has become the shared scoreboard, and Nvidia’s claim to first place among open-weight models refers to that board.
The shift to mixture-of-experts also mirrors the text world. Large total size supports many skills while small active size keeps serving cost down. Speech is borrowing the design that made big text models affordable to run. We expect most future releases in this category to follow it.
Limits: No Streaming and English-Only Audio Questions #
TechCrunch and Tech Times are explicit. Nemotron 3.5 Transcribe handles recorded audio and does not support live streams. Tech Times says real-time streaming is out of scope. That rules out live captions, voice agents and call routing, at least until a streaming variant appears.
The one-hour cap and eight-speaker limit also bind. A long board meeting or a full-day court session needs chunking, and chunking can break speaker identity across segments. Teams should test how labels carry over between pieces before they trust the output.
Speaker attribution covers 13 of the 40 languages. A product that serves all 40 must tell users that diarization is unavailable in the rest. Benchmarks also vary in quality. A 4.62 percent mean across seven languages says little about the other 33.
Five Predictions for the Next Six Months #
These are our analysis, not reported facts. We base them on the release pattern and the limits above.
- A streaming variant is likely. The no-streaming limit is the obvious gap, and ElevenLabs and others are already selling real-time speech.
- Quantized community builds will multiply. With a 17 GB NVFP4 footprint and open weights, hobbyists and small firms will push it toward smaller cards.
- Hosted per-minute prices for batch transcription will fall. Three named hosts now serve the model, so competition among them should trim margins.
- Medical and legal fine-tunes will appear first. The reported drops from 30.8 percent to 11.8 percent WER give vendors a ready pitch.
- Mistral will answer within months. Voxtral Small trails by only 0.04 points on one index, so a refreshed model could retake the open-weight lead.
Frequently Asked Questions #
What is Nemotron 3.5 Transcribe? It is an open-weight speech model from Nvidia, released October 28, 2026. It transcribes 40 languages, translates 31 to or from English, and labels speakers in 13.
How many parameters does it have? Reports differ. MarkTechPost and The Decoder cite about 30 billion total with roughly 4 billion active. Unite.AI and Tech Times cite 34 billion total with about 4.3 billion active. The BF16 memory figure of about 68 GB fits the 34 billion count.
How much GPU memory does it need? Tech Times puts BF16 at about 68 GB, FP8 at about 34 GB and NVFP4 at about 17 GB. Long recordings and batching add memory on top of those figures.
Can it transcribe live audio? No. TechCrunch and Tech Times both say it handles recorded audio only. Real-time streaming is out of scope for this release.
What license does it use? Tech Times and Unite.AI report OpenMDW-1.1. Unite.AI says it allows commercial and non-commercial use. Read the full text on the model page before a commercial launch.
Is it more accurate than Mistral’s Voxtral Small? Not on the Artificial Analysis index. Nemotron scores 2.81 percent WER against 2.77 percent for Voxtral Small, according to The Decoder and Tech Times. Nemotron uses far fewer active parameters.
Where can I download it? Unite.AI lists Hugging Face, GitHub, NGC and Build.Nvidia.com. Tech Times also names ModelScope. Hosted options include Baseten, DeepInfra and Simplismart.