Nemotron-Labs-Audex-30B-A3B is NVIDIA's unified audio-text model, released on July 7, 2026, that packs speech recognition, translation, text-to-speech, and audio understanding into one 30B-total, 3B-active Mixture-of-Experts (MoE) network while claiming to preserve the text intelligence of its backbone. That backbone is Nemotron-Cascade-2-30B-A3B, a text-only MoE LLM; the model supports up to a 1M-token context and ships on Hugging Face under the NVIDIA Oneway Noncommercial License. The crucial claim is that adding audio did not erode text reasoning.

What Shipped: One 30B-A3B Model That Both Listens and Speaks

Audex handles audio understanding, automatic speech recognition (ASR), speech translation, text-to-speech (TTS), audio generation, and speech-to-speech in a single model. Its total parameter count is 30B, but only 3B activate per token, so it aims for 30B-class expressiveness at roughly 3B-class inference cost, the classic MoE trade. A lighter Audex-2B variant was released alongside it.

Architecturally, it bolts an audio path onto an existing text LLM. Per details reported by MarkTechPost, audio enters through an AF-Whisper encoder (16kHz), speech is tokenized by X-Codec2 and general audio by X-Codec, and those tokens flow into the backbone. In other words, sound becomes tokens on the same sequence as text, so the language model reuses the sequence-reasoning ability it already has.

Why "Preserving Text Intelligence" Is the Real Boast

The label "preserves text intelligence" is Audex's most aggressive claim, aimed squarely at a long-standing trade-off in audio multimodal training: teach a language model to hear, and its text-era reasoning and knowledge scores, measured on suites like MMLU-Redux, regress. That the phrase sounds modest only makes the claim bolder. The system gains a new ability at the cost of what it used to do well.

NVIDIA offers benchmarks as evidence that it dodged this trade-off. Among figures compiled by MarkTechPost, the release reports 86.4 on the text-knowledge benchmark MMLU-Redux, a 6.82 word error rate (WER) on OpenASR, and 75.6 on the audio-understanding benchmark MMAU. The core message is that text scores did not collapse relative to the backbone.

This section warrants caution, though. The numbers come from NVIDIA's own materials, and on the Hugging Face model card many metrics are embedded as images rather than readable text. "Did not drop much" is an average impression, not a guarantee of zero loss on every sub-task. Unified-model performance claims always feel different depending on the comparison baseline and task mix.

The License Catch: Open Weights, but No Commercial Use

Do not assume "open source" just because the weights sit on Hugging Face. Audex ships under the NVIDIA Oneway Noncommercial License, a noncommercial license. Research, experimentation, and internal evaluation are open, but commercial use that generates revenue in a product is, in principle, off-limits.

That distinction is decisive in practice. A model under a permissive license like Apache 2.0 can go straight into a startup's service, whereas a noncommercial model, however strong, needs a separate commercial agreement or an alternative before it can reach production. Not conflating strong audio performance with the right to ship it commercially is the first step.

Open Questions for Korean Voice Services

A unified audio LLM paints an appealing picture for Korean call centers, voice assistants, and meeting-transcription services. Merging recognition, translation, and synthesis, once stitched from separate models, into one cuts latency and operational overhead, and a 1M-token context helps handle long calls or lengthy meeting audio in a single pass. The 3B-active MoE design is welcome on serving cost, too.

But the judgment has to be redone in Korean. The published benchmarks center on English-language data, and this material alone cannot tell you the Korean recognition rate, synthesis naturalness, or dialect handling. Add the noncommercial constraint, and for Korean teams Audex reads less as an immediate production candidate than as a reference point for how far unified audio models have come. The direction of binding sound and language into one backbone is now clear; the open question is whether that unification holds under Korean and under commercial terms.

Reference: NVIDIA Releases Audex (Nemotron-Labs-Audex-30B-A3B) (MarkTechPost, 2026-07-07) · Model card (Hugging Face, nvidia/Nemotron-Labs-Audex-30B-A3B)