NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026, according to its Hugging Face model card, which lists the same release date for build.nvidia.com, GitHub and NGC. The card describes it as “an open-weight, offline speech and audio model” for transcription in 40 languages, and says the model “is ready for commercial or non-commercial use” under the OpenMDW-1.1 license.
The card puts the model at 34B total parameters with 4.3B active, including the audio encoder. By our arithmetic, about 12.6% of the parameters are active for each token.
Business Pill · Vertical Integration
An evergreen Business Engineer explainer on vertical integration. It explains the general idea only and does not discuss the companies in this piece.
The key insight: By our arithmetic, Nemotron 3.5 Transcribe runs about 12.6% of its parameters per token, and NVIDIA’s card puts its error rates 4.7% to 28.8% below the next-best open-weight model on three Open ASR Leaderboard tests.
According to the model card, Nemotron 3.5 Transcribe handles recordings up to one hour with a 65,536-token context. It uses a Mamba-2–Transformer hybrid mixture-of-experts language model with a Parallel Expert Encoder that pairs a Transformer ASR encoder with a speaker encoder.
The card lists the tasks as “Transcription in verbatim or written form, contextual transcription, speaker-attributed transcription, rich transcription, speech translation, language identification, and audio question answering”. It adds that the training mixture includes a small amount of PII tagging and redaction tasks.
NVIDIA releases the weights in BF16, FP8 and NVFP4 precision. The card lists deployment on one H100 80 GB GPU and says reasoning can be switched on or off through the chat template.
Mean word error rate on three Open ASR Leaderboard tests (English, long-form, seven multilingual languages) for Nemotron 3.5 Transcribe and the next-best open-weight model named in NVIDIA’s model card. Lower is better. Source: NVIDIA model card, released 28 October 2026, results captured 12 October. Company-reported figures.
The Numbers NVIDIA Published
The card’s results, which it says were “captured on October 12, 2026”, compare the model with the next-best open-weight model on three Open ASR Leaderboard tests. On English, mean word error rate is 4.71 against 4.94 for Qwen3-ASR-1.7B. On long-form audio it is 10.98 against 13.37 for Parakeet-TDT-0.6B-v3. On seven multilingual languages it is 4.62 against 6.49 for Qwen3-ASR-1.7B.
By our arithmetic, those are error rates 4.7%, 17.9% and 28.8% lower than the comparison model on each test. Lower is better on all three measures.
On multi-speaker recordings, the card lists a mean cpWER of 15.80 against 16.83 for a DiCoW v3.3 and DiariZen pipeline, the next-best published result. On style control it lists 7.56 and 6.16 against 8.16 and 8.36 for Canary-Qwen.
Speed, and What Is Not in the Release
The card says one RTX PRO 6000 Blackwell Server Edition GPU, using NVIDIA’s optimized NIM, “transcribes up to 2,238 seconds of audio per second” with NVFP4 and DFlash2 speculative decoding. It adds that “These serving optimizations are not yet part of the open-source release”, and that the open-source vLLM integration “is expected to be roughly two times slower”.
The card also states plainly that “This is an offline model” and not a real-time streaming service.
NVIDIA’s card does not compare Nemotron 3.5 Transcribe with either model, and each company’s figures come from its own tests.
The Structural Read
As we read the model card, NVIDIA is giving away the weights while keeping the fastest serving path in its own NIM: the card says the optimized serving behind its 2,238-seconds-per-second figure is not yet part of the open-source release.
By our arithmetic, the largest relative gap in the card’s open-weight table is on multilingual audio: a mean word error rate of 4.62 against 6.49 for Qwen3-ASR-1.7B, 28.8% lower.
The card names its comparisons as open-weight and published systems; it does not compare the model with Microsoft’s MAI Transcribe or Google’s Gemini 3.5 Transcribe.
NVIDIA, Nemotron 3.5 Transcribe model card, released 28 October 2026
“This model is ready for commercial or non-commercial use.”
Three Implications
FOR DEVELOPERS The card lists deployment on one H100 80 GB GPU, recordings up to one hour and 40 languages, with weights in BF16, FP8 and NVFP4.
FOR TRANSCRIPTION VENDORS The card lists speaker-attributed transcription, speech translation and audio question answering among the tasks of a single open-weight model.
FOR NVIDIA’S HARDWARE The card’s top speed figure is measured on one RTX PRO 6000 Blackwell Server Edition GPU using NVIDIA’s NIM.
The framework states the pattern as: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, NVIDIA opens the model weights while its card ties the fastest serving to NVIDIA’s own GPUs and NIM. The framework does not discuss NVIDIA’s speech models.
What Is Not Established
All benchmark and speed figures come from NVIDIA’s model card; we have not seen an independent evaluation. The card compares the model only with the open-weight and published systems it names, not with Microsoft’s or Google’s models. The percentage reductions and the active-parameter share are our arithmetic, not figures NVIDIA reports.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
NVIDIA released Nemotron 3.5 Transcribe as an open-weight model for commercial or non-commercial use, with 4.3B of 34B parameters active. On the three Open ASR Leaderboard tests in its card, its word error rates are lower than the next-best open-weight model’s, by 4.7% to 28.8% by our arithmetic.
A note on sourcing. Model, benchmark, release and licence details come from NVIDIA’s Nemotron 3.5 Transcribe model card on Hugging Face, read on 29 October 2026. Benchmark and speed figures are company-reported. The active-parameter share and percentage differences are our arithmetic. The Microsoft and Google lines are as recorded in our earlier pieces.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
