NVIDIA opens Nemotron 3.5 Transcribe weights, combining transcription, speaker attribution and translation in one checkpoint
According to an NVIDIA blog dated October 28, 2026, Nemotron 3.5 Transcribe was released as open weights. A single checkpoint covers transcription in 40 languages, speaker attribution in 13 languages, and translation to and from English in 31 languages. NVIDIA also published the Granary v2 data pipeline and a fine-tuning recipe.

What changed
According to a blog NVIDIA published on Hugging Face on October 28, 2026, the company released Nemotron 3.5 Transcribe as open weights. The checkpoint identifier is NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, and it is listed on Hugging Face and ModelScope. It combines transcription, speaker attribution, translation and audio question answering in one checkpoint: it can transcribe 40 languages, translate between 31 of those languages and English, and label speakers in 13 languages. A single recording can be up to one hour long, with up to eight speakers labeled, and word-level timestamps are supported. The architecture is a speech encoder plus a mixture-of-experts language model with a hybrid Mamba–Transformer design, with 30B total parameters and about 4B active per token. The publisher also released the Granary v2 data manifest and NeMo Curator curation pipeline, as well as a NeMo Speech fine-tuning recipe. The blog says the model ranks first on the English open-weight track of the Hugging Face Open ASR leaderboard with 5.33% average WER, with a seven-language multilingual average of 4.62% and 10.98% on long audio. On Artificial Analysis's AA-WER, it is tied for first at 2.8% with another open-weight model, with about 4B active parameters versus the other model's 24B. On David AI's 21-language speaker attribution leaderboard, it ranks first with 18.26% DER, slightly better than NVIDIA's own Nemotron 3 Diarization at 18.38%. In a 14-system comparison by VoiceArena, it achieved the lowest cpWER at 32.4%. These figures are the publisher blog's paraphrase of third-party leaderboard results; leaderboard definitions and snapshot times need independent checking.
What it means for developers
The most direct impact on integration is that a workflow which previously needed a speech recognition model plus a speaker diarization model stitched together can be consolidated into one checkpoint plus a natural-language prompt. The blog describes the mechanism as Sortformer speaker cues and acoustic features entering the language model together, marking turns directly during generation rather than attaching speakers to words after transcription. This is a structural change for overlapping speech and timeline alignment in multi-speaker meetings, not just a stronger recognizer. The capability boundaries are also clear: translation covers only 31 languages with English, not arbitrary language pairs; audio question answering was trained only in English according to the blog; and the model targets recorded audio, not real-time streaming. On deployment, the blog says it can run on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell data-center cards, and provides a vLLM serving path and a NeMo-Speech.cpp local runtime path; hosted endpoints from Baseten, DeepInfra, and Simplismart are also mentioned. The publisher's throughput figure was measured on a single RTX PRO 6000 Blackwell Server Edition at NVFP4 precision, with 30-second clips and high concurrency: 100 hours of audio processed in three minutes. Those conditions differ greatly from real workloads with long recordings, low concurrency, or non-NVFP4 precision. Granary v2 matters because it extends prompt-level control into training: the public manifest covers 48 languages, 16 source corpora, and 463,062 hours, and provides three transcription forms, tn_raw, itn_raw, and itn_clean. The blog explicitly says these numbers describe the dataset release, not Transcribe's full training mix or sampling weights.
Integration and evaluation
The following are judgments and recommendations based on the source, not verified facts. First, decide whether to replace the existing recognition-plus-diarization cascade with a single checkpoint. Replacement is conditional on workloads being mainly recorded audio, clips no longer than one hour, no more than eight speakers, and target languages within coverage. If the product needs real-time captions or streaming transcription, the model is unsuitable by the blog's own description and should be excluded rather than engineered around. The evaluation criterion is to compute both cpWER and DER on your own held-out meeting recordings and compare with the production cascade at the same deletion rate—VoiceArena reported its deleted-word share at 8.32%, versus a comparison median of about 31%; differences in deletion rate can materially affect downstream summarization and retrieval quality. Second, prompt first and fine-tune later: the source's medical case shows that stating the language and writing system in the prompt reduced full-benchmark WER from 30.8% to 23.8%, and fine-tuning reduced it further to 11.8%; on a speaker-split held-out set, it went from 22.44% to 12.17%, a relative reduction of 45.8%. The actionable threshold is therefore: establish a prompt baseline first, and start fine-tuning only if domain entity errors remain high on held-out speakers, provided at least 100 real annotated segments are reserved for checkpoint selection and splits are by speaker rather than random. Evaluate held-out WER, keyword WER, and semantic WER together, plus a general speech regression set—the source itself reports LibriSpeech test-clean rising slightly from 1.54% to 1.69%; general-capability regression must be confirmed on your own corpus and cannot be assumed to be zero. Third, self-hosting versus a hosted endpoint depends on audio volume and data-residency requirements. The evaluation criterion should instead be GPU time consumed per hour of audio under your own clip-length distribution, not the publisher's high-concurrency 30-second-clip figure; at clip lengths near one hour, prefill and VRAM pressure will be entirely different.
Limits and open questions
The blog text calls the model open weights only and does not state license terms in the body; downloadable weights do not equal an open-source license, and commercial use and redistribution rights depend on the model card and accompanying terms. Granary v2 source audio must be downloaded separately under each corpus's original license; the manifest itself does not convey those audio rights. All accuracy and error-rate figures come from the publisher's blog paraphrase of third-party leaderboards and vendor-built cases; leaderboards change over time, the medical case covers only one language and one domain, and the synthetic-data recipe does not guarantee transfer to other industries. Audio question answering trained only in English, translation limited to English pairs, and no streaming support are hard constraints, not things tuning can solve. To be clear: this version of Nemotron 3.5 Transcribe is currently not listed in the KimiSeek model catalog; the Hugging Face and ModelScope downloads and third-party hosted endpoints described in the blog refer to availability from the vendor and its partners, and actual pages and terms should govern before integration.
Illustrations from the original report




