NVIDIA Nemotron 3.5 Transcribe 30B-A4B GGUF

• Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.

---

GGUF builds of nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, NVIDIA's open-weight speech-to-text model (40 languages, recordings up to one hour, word timestamps and speaker labels, speech translation, custom instructions), for NVIDIA's own GGML runtime NeMo-Speech.cpp. By AtomicChat.

A model is two files in one folder: one decoder (pick a size below) and the encoder file NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-Encoder-F16.gguf (speech encoder, speaker branch and CTC timestamp head). The decoder names its encoder file, so keep the encoder's file name as it is.

• Every decoder from 15.55 GB up transcribes like the BF16 decoder on our 2,200-utterance test: English, German, French, Spanish and Chinese. Word error stays within ±0.18 points of BF16 on every English, German, French and Spanish set (table below); Chinese character error moves +0.02 to +0.41 points, every interval including zero, as it does for NVIDIA's own conversions.
• A 13.57 GB decoder for 16 GB GPUs (AD-Q2_0_M): English, German, French and Spanish within noise of BF16, Chinese character error +1.5 points. Peak GPU memory 15.6 GiB with the settings under Run it.
• Smaller than the converter defaults: AD-NVFP4 is 19.68 GB on disk against 21.79 GB for NVIDIA's NVFP4 conversion, AD-IQ4_NL 19.68 GB against 20.35 GB for its IQ4_NL conversion, same expert formats; the embeddings and output head are Q8_0 instead of BF16, and the IQ4_NL experts are quantized with an importance matrix.
• Seven decoders, 13.57 to 35.00 GB on disk, from Blackwell NVFP4 to IQ4_NL and Q8_0 for CPU, Metal and Vulkan.
• Ready-made files: no 67 GB checkpoint download and no Python conversion.
• DFlash2 drafter included (BF16, 1.15 GB) for speculative decoding on the batched CUDA path.

Files

Word error against the BF16 decoder

Change in word error (character error for Chinese) against the BF16 decoder, in points, with the 95% interval from a paired bootstrap over utterances. Negative is fewer errors. Rows in italics are not in this repo: the BF16 reference and NVIDIA's converter defaults (convert_model.py --outtype nvfp4 / iq4_nl), measured the same way. The last column counts the 2,200 utterances whose transcript is word for word the BF16 decoder's; it is noisy (batching order changes it: the same AD-IQ4_NL file run on a second box matched its own first run on 1987), so differences of a few dozen mean nothing.

Test set (fixed sample, seed 0): LibriSpeech test-clean 500 and test-other 500, Multilingual LibriSpeech test German, French and Spanish 300 each, FLEURS test Mandarin 300; 6.7 hours of audio. Greedy decoding through nemo-speech transcribe (batched CUDA path, language hint per set) on an A100 80 GB, NeMo-Speech.cpp 6a62501 (llama.cpp bd4f514). Whisper's English normalizer for English, its basic normalizer for the rest. These are relative numbers on our sample, not leaderboard scores.

¹ Measured on a second box (A100 40 GB, same build). Moving boxes alone shifts a file by up to ±0.17 points: the same AD-IQ4_NL file read -0.17 [-0.36, -0.01] on MLS German there.

Run it

Build NeMo-Speech.cpp (build guide): CMake 3.26+, SentencePiece 0.2.0+ (Ubuntu 22.04's packaged 0.1.96 fails to compile), CUDA 12.8+ for the batched CUDA path.

git clone https://github.com/NVIDIA/NeMo-Speech.cpp && cd NeMo-Speech.cpp
git submodule update --init --depth 1 llama.cpp third_party/cpp-httplib
scripts/configure.sh cuda-server && cmake --build --preset cuda-server

hf download AtomicChat/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-GGUF --local-dir models \
  --include "*Encoder-F16.gguf" --include "*AD-NVFP4.gguf"

# one file, JSON with word timestamps
build/cuda-server/bin/nemo-speech transcribe audio.wav \
  --model models/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-AD-NVFP4.gguf --format json --word-times

# OpenAI-compatible server: /v1/audio/transcriptions, /v1/chat/completions
build/cuda-server/bin/nemo-speech serve --asr-model models/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-AD-NVFP4.gguf --port 8080

• Speaker labels: --diarize. Translation: --translate-to de. Custom prompt: --instruction "...".
• Smaller GPUs: lower --asr.salm.max_sequences and raise --asr.salm.n_ctx_per_sequence; NVIDIA's configuration guide lists the settings per GPU size.
• 16 GB GPUs with AD-Q2_0_M: --asr.salm.max_sequences 2 --asr.salm.n_ctx_per_sequence 8192 --asr.salm.max_context 8192 --asr.salm.mtp_draft 0 --asr.salm.n_batch 512 --asr.salm.encoder_max_batch 2. Peak GPU memory 15.6 GiB on a 3.7-minute file (measured on an A100, not on a 16 GB card: it is tight). max_context 8192 holds about 7 minutes of audio per request.
• DFlash2: --asr.salm.dflash_path models/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-DFlash2-BF16.gguf (batched CUDA path).
• CPU, Metal and Vulkan builds use upstream llama.cpp kernels: take AD-IQ4_NL (NVIDIA lists IQ4_NL for these paths) or Q8_0. Our measurements ran on CUDA only.

How the files were made

• NVIDIA's converter (convert_model.py --architecture salm) wrote the encoder file and a BF16 decoder; the decoder is a standard llama.cpp nemotron_h_moe GGUF with the nemo_speech.* keys that pair it with the encoder.
• Rungs were quantized from that BF16 decoder with the converter's own llama-quantize. Experts carry the size; attention, Mamba in/out, the shared expert, the embeddings and the output head are Q8_0; the MTP head keeps the experts' type. Most matrices have rows of 2688, 1856 or 3712, which only block-32 types fit (no K- or I-quants with 256-blocks), so these files use NVFP4, Q8_0 and other block-32 types; the tensor types inside each file are the ones its name says.
• AD-IQ4_NL, AD-IQ4_NL_L, AD-Q5_0_XL and the two AD-Q2_0 files use an importance matrix from the BF16 decoder over 1.5M tokens of our text calibration corpus (imatrix.gguf here). It is text only: llama-imatrix cannot feed audio, so the activations on audio positions are not in it, which is why every file was checked on speech. At 4 bits it buys nothing measurable: NVIDIA's IQ4_NL conversion without one is within noise too. Q8_0 and AD-NVFP4 read no importance matrix (neither format uses one).
• Below 4 bits the quantizer decides everything. Stock Q2_0 sets the block scale to the block's largest weight, so it writes three levels, sends most weights to zero and ignores the importance matrix: with it on every expert (11.06 GB) English word error rose by 3.5 points and German, French and Spanish by 14 to 25. The AD-Q2_0 files here use our imatrix-weighted Q2_0 quantizer instead (all four levels, scale fitted per block; same Q2_0 format, so the stock runtime reads it). AD-Q2_0_M keeps the 6 MoE blocks with the largest activations (40-51, from the importance matrix) and the MTP head at IQ4_NL; AD-Q2_0_L puts only the experts' down-projections at Q2_0 (down-projections at Q2_0 beat up-projections at Q2_0 on German and Chinese at the same size). The same quantizer with every expert at Q2_0 (11.06 GB) still costs +0.2 points English and +0.8 German, +2.2 Chinese character error, so it is not published.

License

The model is under the OpenMDW License 1.1 (LICENSE in this repo), as is its base, NVIDIA-Nemotron-3.5-Lightning-30B-A3B. These are conversions; see NVIDIA's model card for the training data, evaluation, intended use and limitations.