# NVIDIA Nemotron Halloween Project Hits 0.72s Latency

URL: https://3ammarketer.com/news/nvidia-nemotron-halloween-model
Published: 2026-10-31

> NVIDIA detailed its MonsterHouse interactive installation, showing how Nemotron 3.5 Bolt E4B and NIM components achieve sub-second conversational latency on local hardware.

The NVIDIA MonsterHouse interactive Halloween project achieves a median end-of-speech-to-audio latency of 1.70 seconds for standard replies and 0.72 seconds for cached responses. Coordinated by the NVIDIA Nemotron Voice Agent, the real-time installation processes voice recognition, character reasoning, emotional speech generation, and facial rendering entirely on local hardware. [1]

## Workstation architecture and character rendering

The entire system runs on a single workstation equipped with an NVIDIA RTX PRO 6000 Blackwell Max-Q and an NVIDIA RTX PRO 4000 Blackwell. For the visual front end, the character SkelMesh was built in Autodesk Maya and exported to Unreal Engine 5, where real-time virtual fire and visual effects render behind the character door. This setup provides an operational blueprint for deploying [autonomous AI agents](/blog/autonomous-ai-agents-runtimes-costs) and rich 3D graphics simultaneously without remote cloud hops. [1]

## Nemotron 3.5 Bolt E4B reasoning engine

To handle dialogue, the project deploys NVIDIA Nemotron 3.5 Bolt E4B 4GB, an open small language model designed for conversational workflows. The project uses the 4-bit quantized release by default, though a BF16 checkpoint remains available for comparison. The project's author noted that Bolt E4B carries a smaller footprint than alternative frontrunners like Qwen3 30B-A3B and Qwen 3.6 35B-A3B. To shape the character personality, the author applied a LoRA adapter trained on a dataset of over 5,000 multi-turn conversations. [1]

From our data at 3AM Marketer tracking open foundation weights, NVIDIA released NV-Reason-CT on 2026-09-08 with 5.3B parameters under the openmdw-1.1 licence, alongside the massive 560.5B-parameter Nemotron-3-Labs-Ultra-Math-RL on 2026-09-02. [2][3]

## Audio perception and Audio2Face-3D animation

Perception and voice components operate in streaming tandem. Visitor speech is captured by NVIDIA Nemotron 3 ASR via the nemotron-speech-streaming-en-0.6b.q8_0.gguf model checkpoint, with speech-end detection configured to trigger after 300 ms of silence. Environmental vision tasks rely on NVIDIA Nemotron 3 Nano Omni to describe visitor costumes, running alongside YOLO and ByteTrack for crowd counting. [1]

For the return audio path, NVIDIA Magpie TTS Zeroshot synthesizes spoken replies from short reference voice clips. Because each distinct vocal emotion requires a reference sample of 3 to 10 seconds, the interface incorporates a dedicated voice sampling control block. Clean synthetic audio feeds directly into NVIDIA Audio2Face-3D, which generates procedural mouth animations inside Unreal Engine 5 in real time. [1]

Our read: A 4-bit small model on one workstation can run a voice character with no cloud call. The bigger lever is caching: cached replies came back in 0.72 seconds, well under generated ones, so scripting common lines is worth more than a faster model.

Limit: The reported 1.70-second and 0.72-second latencies reflect a dual-GPU Blackwell workstation with local streaming models. Environments using lower-tier graphics silicon, unquantized BF16 weights, or longer speech-end silence thresholds will see higher latency.

## Sources

1. [nemotron-halloween](https://huggingface.co/blog/nvidia/nemotron-halloween)
2. [Open model release tracker (3AM Marketer data set)](https://3ammarketer.com/data/open-models)
3. [llama-3.1-nemotron-70b-instruct Model by NVIDIA | NVIDIA NIM](https://build.nvidia.com/nvidia/llama-3_1-nemotron-70b-instruct)