NVIDIA's open voice-agent blueprint shipped two new small models and a conversational-mode toggle
AspireAI community · GitHub / NVIDIA
Nemotron Voice Agent v2.3.1, released 29 October 2026, adds Nemotron 3.5 Bolt 74B A9B on single-GPU hosts and Nemotron 3.5 Bolt E4B on NVIDIA Cloud, server and single-GPU hosts. The 74B A9B is served with vLLM on a Blackwell workstation or DGX Spark, with LiLiCorr speculative decoding and streaming input; the E4B comes as a NIM on server recipes and with vLLM on single-GPU recipes, using DFlash speculative decoding on Blackwell GPUs, and its services card gains a Conversational Mode toggle that is on by default - the same speech-ready short-response behaviour NVIDIA describes on the model card, where conversational mode exists to encourage concise answers for natural back-and-forth. The release also changes reasoning behaviour in a way that will bite anyone scripting against it: when Max Tokens is set, an LLM's Reasoning Budget stays at 64 tokens under it, and under 192 Max Tokens the backend turns Reasoning off for the session. The blueprint itself is a reference implementation on the open Pipecat framework that chains Nemotron or Parakeet ASR, a Nemotron LLM and Magpie or Chatterbox TTS, with voice activity detection and end-of-utterance handling for interruption, an OpenAI Realtime-compatible WebSocket gateway, and example pipelines for generic, multilingual, Omni, Omni-with-subagents and frontend/backend-agent patterns. The repository is dual-licensed, with source under Apache-2.0.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.