Open Model WeightsNVIDIA · FIELD VERIFIED
Primary sources last checked on 30 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records NVIDIA Nemotron 3.5 Transcribe 30B-A4B as a 34B total / 4.3B active (publisher-declared)-parameter audio and text instructions → text and word timestamps open-weight model in the Nemotron 3.5 Transcribe family, with NVIDIA listed as the developer. The checked source repository provides 1 recognized weight artifacts, format evidence for Safetensors and a verified context window of 262,144 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares NVIDIA OpenMDW License Agreement v1.1. NVIDIA permits commercial and noncommercial model use subject to OpenMDW-1.1, redistribution notices and applicable component terms.. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
The checked sources declare nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 as lineage evidence. Training recipe status: training-described-in-model-card; data disclosure status: dataset-metadata-declared.
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-30
NVIDIA Nemotron 3.5 Transcribe 30B-A4B is an officially released open-weight offline speech and audio model combining a speech encoder with a sparse language backbone. NVIDIA reports 34 billion total parameters including audio components, approximately 4.3 billion active, support for 40 languages and recordings up to one hour. The source-verified repository declares OpenMDW License Agreement v1.1. Open Model Weights classifies the original BF16 checkpoint as Notable for speech transcription, translation and speaker-attributed word timestamps; publisher WER or throughput reports are not independent OMW tests.
WHY IT MATTERS
The release is a distinct, reproducible speech-specialist checkpoint rather than another generic chat model or a runtime-only feature announcement.
Publisher-described tasks include verbatim and normalized transcripts, prompted terms, translation, language identification, speaker diarization and timestamp-rich output.
An official downloadable checkpoint, dedicated technical explanation and explicit conditional OpenMDW-1.1 grant make the release useful for informed offline speech deployments.
MODEL-SELECTION VIEW
NVIDIA describes speech instruction-following tasks including verbatim ASR, normalized transcripts, language identification, cross-language translation, audio question answering and diarized transcripts with word timestamps. A 40-language coverage claim needs real-language and noise-conditioned evaluation, not a generic guarantee of equal recognition quality.
The released BF16 model's repository is the source for weights and configuration. Additional optimized variants and speculative drafts should be identified by separate checkpoint IDs instead of being falsely presented as new independent model architectures.
The system contains an audio front end and a sparse language component. NVIDIA's 34B total and approximately 4.3B active counts are publisher architecture claims, not a measurement that the complete model fits in 4.3B-equivalent weight memory.
The NVIDIA research description says offline processing covers up to one-hour recordings. Operational throughput and peak memory depend on encoder preprocessing, windowing, speaker attribution, batching, runtime optimization and device capacity.
NVIDIA reports mean WER of 4.71% on an English Open ASR collection and 4.62% on the multilingual selection, with separate analyses of overlapping speakers and named entities. The numbers are useful provenance, but remain publisher-supplied.
Some high-throughput NIM-serving results rely on performance optimizations not yet part of the public release. Open Model Weights has not run the checkpoint, replicated WER, validated speaker diarization or measured streaming behavior.
Notable highlights specialist relevance with downloadable official weights and source-described evaluation scope. This checkpoint addresses speech-to-text quality, not general-purpose language intelligence ranking.
Without a reviewed independent, checkpoint-equivalent evaluation against peer offline speech models, Leading would overstate the evidence. The original license and file manifest remain more important than a promotional rank.
The published BF16 source checkpoint is the reference for this entry; FP8, NVFP4, GGUF and draft components are distinct distributions, each with an exact model identifier and their own runtime prerequisites. The audio encoder, speaker-information network and text processor affect the end-to-end data path.
A meaningful deployment audit must record resampling, channel count, signal duration, language prompts, speaker hints, chunk treatment and timestamp alignment along with the model revision. Small pipeline differences can substantially change word-error and diarization results without changing the underlying source model.
Forty-language support is a declared model scope, not a promise of uniform accuracy across accents, noisy rooms, dialects or specialist vocabularies. Build separate language- and speaker-stratified test sets, retain reference transcriptions, calculate error metrics against the intended task and review time-aligned outputs manually.
Speech recordings often contain third-party personal data. Consent, retention, geographic processing, access control and human confirmation of quoted words matter as much as model choice. Offline processing can help manage data locality, but the way operators actually capture, store and route audio still determines privacy risk.
IMPLEMENTATION & EVALUATION
A reproducible offline ASR evaluation plan; these steps are not claims that Open Model Weights ran the model.
Audio systems may require separate encoder, tokenizer and task templates.
Record exact repository SHA, weight manifest, preprocessing parameters, code revision, language settings and output schema.
Counting third-party quantized weights as the original or omitting audio preprocessing dependencies.
Recognition errors vary with accent, language, noise, overlap and vocabulary.
Score task-specific WER/CER, timing accuracy and diarization per test subset, separately from translated output quality.
Applying a single English WER to every speaker and language or treating time-aligned word text as guaranteed.
Real-world recorded speech introduces personal-data and serving constraints.
Verify OpenMDW-1.1 notices, consent/retention requirements, actual peak memory, CPU/GPU use and end-to-end latency.
Mistaking a publisher-optimized NIM benchmark for the public weights or streaming inference performance.
Serving-optimized benchmark results may require components not released with the checkpoint.
Separate attributed NVIDIA WER and NIM claims from any measured local runs, with full hardware and audio-test conditions documented.
Reporting publisher-only metrics as independently observed or assuming non-upstreamed speedups exist in the public weights.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
No. This original checkpoint is built for speech and audio-conditioned instruction processing. Its Notable selection is specialist relevance, not a general intelligence ranking.
NVIDIA describes an offline system for recordings up to one hour, not a validated real-time streaming ASR API. Local production streaming capability has not been verified by Open Model Weights.
No. WER claims are attributed to NVIDIA; Open Model Weights has not run a matched benchmark.
NVIDIA's OpenMDW License Agreement v1.1 includes conditional commercial and noncommercial permission. Read the current license and notices for the exact model and intended use.
No. The 34B count includes the audio-related components, while active parameters describe a sparse model's computation on a token. Neither figure alone determines downloaded checkpoint bytes, GPU memory or serving throughput.
No. Our Notable designation reflects a distinct, verifiable specialist release. Comparative leadership would require an independent speech evaluation of matching public checkpoints, tasks and datasets, which Open Model Weights has not performed.
03 · PRACTICAL TECHNICAL PROFILE
Inspect weight-memory estimates, model lineage and runtime evidence separately from promotional or untested performance claims.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
NVIDIA permits commercial and noncommercial model use subject to OpenMDW-1.1, redistribution notices and applicable component terms.
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
04 · EVIDENCE AND UPDATES
Inspect dated source checks and any differences between published claims and the evidence; these checks are not independent model benchmarks.
SOURCE-BASED VISUAL BRIEF · NOTABLE
Three evidence-linked visual summaries: parameter scale, illustrative precision memory and verifiable provenance. No performance ranking is implied.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
05 · SOURCE RECORD
Full repository-level attributes, licensing, format details and provenance are available below without hiding or rewriting the original data.
01 · WEIGHTS
02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
License / model-card evidence ↗03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: NVIDIA OpenMDW License Agreement v1.1 (allowed-with-conditions commercial-use classification; see original terms).
Context: 262144 configured tokens.
BF16 weights: 68.0 estimated decimal GB, not runtime VRAM.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →