NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026 on Hugging Face, NGC and build.nvidia.com, under the OpenMDW-1.1 license. It is an offline speech-augmented language model. A plain-language instruction selects the task: transcription, speaker-attributed transcription, translation to and from English, timestamps, language identification or audio question answering. One recording can be up to an hour long, within a 65,536-token context.
What the model card claims
The architecture is a Mamba-2 and Transformer hybrid mixture of experts with an audio encoder and a speaker encoder. The card lists 34B total parameters, 4.3B active, both including the audio encoder. Weights ship in BF16, FP8 and NVFP4.
The card's benchmark results, captured October 12, include a mean WER of 4.71 on the nine English sets of the Open ASR Leaderboard, against 4.94 for the next-best open-weight model, and a mean cpWER of 15.80 on multi-speaker full sessions. These are vendor-reported numbers. NVIDIA's launch blog says the model "ranks No. 1 on the Hugging Face Open ASR" leaderboard.
Throughput claims need care. The card says one RTX PRO 6000 Blackwell Server Edition GPU transcribes "up to 2,238 seconds of audio per second" with NVFP4 and speculative decoding, but that figure comes from the Nemotron Transcribe NIM. The card adds that "these serving optimizations are not yet part of the open-source release" and that the open vLLM path is "expected to be roughly two times slower." For batched English ASR on one H100 in BF16, the card gives an approximate RTFx of 260 to 410, depending on the speculative decoding mode.
Language coverage, checked against the matrix
The launch blog says the model covers 40 languages, "including seven Indic languages." The card's language matrix lists five Indian languages. Hindi is rated "Leading". Bengali, Marathi, Tamil and Telugu are rated "Supported", which the card defines as "may need fine-tuning", and the matrix does not specify translation for any of the four. English includes an India regional variant. The card reports Hindi–English and Bengali–English code-switching results on MUCS 2021, at 19.28 and 35.46 percent WER respectively. Teams building for Indian languages should treat anything beyond Hindi as a starting point for fine-tuning, not a finished product.
Why active parameters do not decide the GPU
A mixture of experts runs only some of its weights for each token, which cuts compute. But every expert must still sit in memory. Here is a rough calculation of weight memory alone, ignoring activations, recurrent state and the runtime's own overhead:
- BF16 at 2 bytes per parameter: 34B × 2 ≈ 68 GB.
- FP8 at 1 byte: ≈ 34 GB.
- NVFP4 at 4 bits, plus one 8-bit scale per 16 values (about 4.5 bits each): ≈ 19 GB.
The card's hardware table matches this calculation. A 32 GB RTX 5090 is listed for NVFP4 only. A 48 GB L40S is listed for FP8. BF16 appears only on 80 GB and larger parts. The 4.3B active figure explains why the model is fast once loaded. It does not shrink what must be loaded. The launch blog also says quantized versions are "coming soon" for RTX Spark laptops, while the card already links FP8 and NVFP4 checkpoints. RTX Spark is not in the card's hardware table.
What remains uncertain. The benchmarks are NVIDIA's, and the fastest serving path is not open source yet. The card warns that audio understanding and spoken instruction following "were evaluated only in English." It is an offline model, not a streaming service.
Analysis: When sizing any mixture-of-experts model, multiply total parameters by bytes per weight before you look at the active count. For a help desk like the one at Belgavi's Book Store, where Geetha wants call recordings turned into case notes with speaker labels, the practical choice is the NVFP4 checkpoint on a single 32 GB card or FP8 on a 48 GB card. Measure WER on your own recordings before trusting the leaderboard figure.
Sources and evidence
NVIDIA's open-weight Nemotron 3.5 Transcribe (October 28) is a 34B-total, 4.3B-active hybrid MoE speech model for transcription, speaker labels, translation and audio Q&A over recordings up to an hour; its card lists five Indian languages, not the blog's seven, notes the fastest serving numbers come from a NIM that is not yet open source, and its hardware table matches a simple memory calculation of about 68 GB BF16, 34 GB FP8 and 19 GB NVFP4.