The Pulse
NVIDIA updates Nemotron 3.5 Transcribe GGUF decoders
NVIDIA's GGUF repository now includes an IQ4_NL decoder and an NVFP4 decoder converted from the NVFP4 checkpoint. The offline speech model supports transcription in 40 languages and recordings up to one hour.

NVIDIA updated its Nemotron 3.5 Transcribe GGUF package with an IQ4_NL decoder and an NVFP4 decoder converted from the NVFP4 checkpoint on November 18, according to its Hugging Face commit history. The open-weight, offline model supports transcription in 40 languages and recordings up to one hour.
The GGUF files are intended for NeMo-Speech.cpp, NVIDIA's native C++ runtime for the Nemotron Speech model family. The runtime supports Linux and Windows, and the model is available for commercial and non-commercial use.
Two files for a model
A working model consists of a language-model decoder and a shared encoder file. The encoder contains the audio encoder, its speaker branch and a CTC timestamp head; the runtime loads it from the same directory as the selected decoder.
The NVFP4 decoder is 21.2 GiB and is intended for Blackwell GPUs. The IQ4_NL decoder is 19.0 GiB and is intended for other GPUs, including Ampere, Ada and Hopper, as well as CPUs. The encoder file is 1.6 GiB and is required with either decoder.
For NVFP4, the listed hardware requirement is an NVIDIA Blackwell GPU with 32 GB of memory or more. For IQ4_NL, the card lists other NVIDIA GPUs with 24 GB or more, or a CPU. Its integration table specifies settings from the runtime guide for the 24 GB RTX 3090 and RTX 4090.
Speech tasks and runtime
The model processes recordings alongside text instructions and optional context. Its supported tasks include multilingual transcription, speech translation, speaker-attributed transcription and answering questions about audio. It can also return word timestamps and speaker labels.
The model has 34 billion parameters, including the audio encoder, of which 4.3 billion are active. Its language model uses a Mamba-2–Transformer hybrid mixture-of-experts architecture, paired with a Parallel Expert Encoder.
NeMo-Speech.cpp can transcribe from the command line or run an OpenAI-compatible server. The server can return word timestamps and speaker labels, and offers subtitle output in SRT or VTT format. Applications must supply conversation history; the model recipe does not establish persistent memory.
Decoding and evaluation
On CUDA, the decoder's multi-token prediction head drafts seven tokens per step by default for the model to verify in one pass. The draft count can be set up to 15, and setting it to zero turns the feature off.
An optional DFlash2 draft model is available as a 1.1 GiB file. NVIDIA says it provides faster decoding when there are few requests at once. The model card points to benchmark results for the base model but notes that those results were computed for the BF16 model, not the GGUF variants.
The model is governed by the OpenMDW-1.1 license. NVIDIA says integration into AI systems requires additional testing with use-case-specific data.
Sources
- Original source NVIDIA · NVIDIA Nemotron 3.5 Transcribe 30B A4B GGUF huggingface.co
Other pages read
- NVIDIA · page 2 huggingface.co
- Hugging Face huggingface.co
In the tracker
Related stories
- Salesforce releases open-weight speech transcription model
- Nvidia releases chemistry dataset with Nemotron for Science
- Reson8 releases open speech model for more than 40 languages
- ElevenLabs releases Scribe v3 transcription models, citing fewer errors
- NVIDIA adds packet reordering to DAQIRI data acquisition library