Summary
Fahd Mirza installs and tests NVIDIA’s Audex 2B, a newly released 2-billion-parameter model from the Neotron family that unifies audio understanding and generation with standard text reasoning in a single model. Capabilities demonstrated include speech transcription, multi-language translation (Spanish, Russian), long-form audio summarization, text-to-speech synthesis, and general text reasoning — all within one model that consumes roughly 6 GB of VRAM on an NVIDIA RTX 6000.
The architecture explanation is a highlight: Audex 2B uses a standard transformer LLM as its reasoning core, bolting on an audio encoder (MLP adapters convert raw audio into embeddings the LLM can process) and separate discrete speech/audio token decoders that convert outputs back into audio waveforms. Training follows a two-phase pipeline — supervised fine-tuning using both multi-stage curriculum and consolidated single-stage approaches, followed by cascaded reinforcement learning (the same technique used in Neotron Cascade 2), with a final MOPD distillation step to compress reasoning into a faster model. The model supports a 128k context window and runs in both thinking and non-thinking modes.
Running at approximately 11 tokens per second on a single GPU, Mirza puts the model through practical tasks including transcribing a Napoleon historical narration, summarizing an 11-minute audio file in five sentences, and translating speech clips. Results are generally strong, with minor edge cases (clipping on the final word of a short transcription). NVIDIA’s all-inclusive setup script from Hugging Face and the Gradio interface make local deployment accessible, and Mirza provides a GPU rental coupon for those without local hardware.
📺 Source: Fahd Mirza · Published July 08, 2026
🏷️ Format: Tutorial Demo
