# Deepgram Targets Air-Gapped Laptops for Voice AI Models

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/30535
Published: 2026-10-12T19:24:30.000Z
Updated: 2026-10-12T19:24:30.000Z
Section: Technology

> The company already supports self-hosted deployments, including an air-gapped Amazon SageMaker setup, but has not announced native laptop availability.

Deepgram is pursuing voice-AI models that run directly on customer hardware, including offline laptops, a move that would extend its self-hosted technology beyond cloud and data-center environments.

Nick Mann, Deepgram’s staff technical program manager, said Monday that the company wants its models to operate natively on customer devices without an internet connection.

“I want our models on those, on the hardware, to run natively. Air-gapped. That’s where we want to get to,” Mann said.

Deepgram has not announced native laptop deployment or provided a launch date, supported hardware, operating system or model size. Its current self-hosted deployments can run on customer-requisitioned cloud infrastructure, bare metal or customer data centers.

The company also supports a fully air-gapped deployment using Amazon SageMaker. In that configuration, the deployment is isolated from Deepgram Cloud, allowing customers to keep audio and related processing within their controlled environment.

Deepgram positions self-hosting for customers with performance, security and data-privacy requirements. In a typical self-hosted deployment, audio, transcripts and identifying request markers are not sent to Deepgram.

The company announced general availability of Flux Multilingual on April 29, 2026. The speech-recognition model supports English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch, and is available through Deepgram’s cloud API or as a self-hosted deployment.

Flux is designed for conversational speech recognition, including turn detection and interruption handling. Deepgram also develops speech-to-text, text-to-speech and voice-agent technology.

Mann said Deepgram has 45 researchers and that developing its models requires substantial computing power and data.

“You know what they like to eat? Compute. Just tons of compute, tons of data,” he said.

Mann also said a large percentage of Deepgram customers host the company’s models themselves, without specifying a share or customer count. He raised the possibility of retraining the models with improved data, asking, “What’s stopping us from retraining it on better data?”