Nemotron 3.5 Transcribe is 34B parameters under a licence that is not Apache

NVIDIA's announcement on October 28 describes an open speech model that transcribes 40 languages, translates 31 of them, labels up to eight speakers in one pass and answers questions about a recording. All of that is broadly accurate for the checkpoint now sitting on Hugging Face. What it skips is the housekeeping, which is the interesting part: the main repository was created on October 7, three weeks before anyone called it a release, and the two quantised variants the card links to return an authentication error instead of weights.

What is actually downloadable

The collection stamped 15:58 UTC on October 28 holds three items: the BF16 checkpoint, a small draft model for speculative decoding, and a demo Space. All three existed earlier, which makes the announcement a publication event rather than a creation event.

Repository	Created (UTC)	Downloads	What it is
NVIDIA-Nemotron-3.5-Transcribe-30B-A4B	7 Oct, 22:50	338	33.7B parameters, single BF16 shard
...-30B-A4B-DFlash2	7 Oct, 23:35	37	0.57B draft model, no word timestamps
FP8 and NVFP4 variants	not public	—	linked from the card, return 401

The blog says quantised builds are "coming soon", which matches what the pages do. The practical consequence: the only widely runnable artefact wants one H100 80 GB at BF16, through vLLM 0.32.0 and the NeMo Speech main branch. Fine-tuning tooling ships separately. A 338-download model is not a deployed model; the same organisation's 0.6B streaming checkpoint from May has passed 923,000.

Inside a 34-billion-parameter speech model

The checkpoint is a Speech-Augmented Language Model: a Mamba-2 and Transformer hybrid mixture-of-experts backbone — the 30B-A3B text model from the Bolt generation — wrapped in roughly 800M parameters of audio perception — a 600M ASR transformer, a 99M Sortformer speaker encoder and a CTC alignment head that produces word-level timestamps. Active compute is 4.3B parameters, not the 4B the trade press reported; the total is 34B, not 30B.

Context is 65,536 tokens; the supported recording length is one hour. SiliconANGLE wrote that it digests "100 hours of audio"; the card says recordings up to one hour and warns the model is not guaranteed to generalise past either limit. A tenfold error in the stated ceiling is what an evaluation team finds in week two.

Coverage is less uniform than the headline. The matrix lists 40 locales, marks translation on 32 while the blog says 31, and restricts speaker attribution to 13. Norwegian appears as supported with every capability column blank.

The numbers are NVIDIA's own

Every figure below comes from the model card, captured October 12, with no independent replication yet. The leaderboard submission filed in the repository attributes its scores to the model card itself, which is self-reporting wearing a leaderboard's clothes.

Benchmark	Metric	Nemotron 3.5 Transcribe	Closest open-weight rival
Open ASR Leaderboard, 9 English sets	mean WER	4.71	Qwen3-ASR-1.7B, 4.94
Open ASR Leaderboard, long-form	mean WER	10.98	Parakeet-TDT-0.6B-v3, 13.37
Multilingual Open ASR, 7 languages	mean WER	4.62	Qwen3-ASR-1.7B, 6.49
AMI, CALLHOME, Mixer 6, NOTSOFAR-1	mean cpWER	15.80	DiCoW v3.3 plus DiariZen, 16.83

Averages hide the per-locale spread that decides usability in a given market. On FLEURS test sets the card reports English at 3.35 WER, French at 3.16, Danish at 9.11, Filipino at 18.28, and Chinese at 8.62 CER.

Efficiency claims carry the same qualification. One RTX PRO 6000 Blackwell Server Edition is credited with 2,238 seconds of audio per second and a 0.13-second transcript of a 30-second clip, measured on the NIM build with NVFP4 and speculative decoding. The next sentence says those optimisations are not in the open-source release and that the vLLM integration shipping on release day should run roughly twice as slow. The best number and the downloadable number are different numbers.

Granary v2 is the more reusable artefact

The dataset is where the release earns its keep. Granary v2 covers 48 languages across 16 source corpora, with labels for 463,062 hours of audio, 165,300 of those flagged high-quality by the pipeline's own gate, 92,496 hours of speech translation into English and 39,375 hours out of English fanned across 31 target languages. Each utterance carries six text targets: normalised transcript, inverse text normalisation, captioning, code-switched text, speech QA and translation.

Granary v1, from May 2025, was 25 European languages and roughly a million hours of pseudo-labelled speech, two tasks, a Whisper-large-v3 pipeline. The version number understates the change: v2 adds the instruction layer, the translation fan-out, contextual biasing with distractor terms and entity annotations, and rebuilds the labelling stack in NeMo Curator, whose pipeline code is already public.

Two caveats decide its usefulness. The repository holds label manifests, not audio — you fetch VoxPopuli, LibriLight, Common Voice and the rest from their distributors under their own terms, and the largest single block is 380,711 hours of VoxPopuli. And the labels are OpenMDW-1.1 except for the People's Speech subset, which is CC BY-SA 4.0. NVIDIA also says the full end-to-end runs and per-stage scripts will follow soon, which is a promise, not a deliverable.

For fine-tuning that is a bigger deal than any leaderboard row. NVIDIA's medical case study reports a 45.8% relative WER cut on held-out speakers, but it is a vendor-run experiment with no released results to check.

The predecessor nobody mentions

The model people were already using is nvidia/nemotron-3.5-asr-streaming-0.6b, a 600M FastConformer-RNNT checkpoint from May with configurable chunks of 80, 160, 320, 560 and 1,120 milliseconds. It does not translate, does not label speakers and does not answer questions. It streams.

That distinction got lost in the coverage. The "sub-100 ms" figure attached to the family comes from the smallest chunk setting on the streaming model, and Transcribe is explicitly an offline model — the card even states that its 30-second encoder chunking is "not a streaming latency claim". October 28 added breadth: translation, diarization, prompting, timestamps, audio QA, a much larger backbone. It did not add a replacement for the voice-agent use case the 0.6B model serves.

Why an open ASR stack is a pricing argument

Speech APIs bill by the minute; a checkpoint with 40 locales and diarization is a credible reason to stop paying that. The licence is where the argument gets complicated. The weights ship under OpenMDW-1.1, which Hugging Face tags as "other", not Apache 2.0. The card says commercial use is permitted, and the agreement is a model-distribution licence with attribution and notice obligations rather than the patent grant and trademark permission of the Apache mould. Legal teams read that difference; nobody else does.

The other boring question: do the WER numbers travel? A 4.71 English mean assembled from LibriSpeech, SPGISpeech and Earnings22 says little about a call centre in Jakarta, and the card's own FLEURS spread runs from 3.16 to 18.28. Run your own audio through the demo before believing the average — and note that the H100 requirement, not the licence, is the barrier to running this on-premises today.