Elevenlabs Hosted Llms

A voice agent chains several models: one turns the caller's speech into text, a language model decides what to say or which tool to call, and another turns the reply back into speech. In this November 5, 2025 post, ElevenLabs says it began running open-weight language models on its own infrastructure inside its Agents Platform, alongside its speech-to-text, text-to-speech and turn-taking models. The point is that customers can build a voice agent without sending each turn to a separate LLM provider, and ElevenLabs argues that keeping every model in one environment reduces latency and improves reliability and data security.

The post names two launch models. ElevenLabs says GLM 4.5 Air, from Z.ai, delivers top-tier reasoning and tool-calling accuracy at roughly one-third the cost of alternatives, and that Qwen's Qwen3-30b-a3b, meant for lighter reasoning, reaches a time to first sentence under 150 milliseconds. These are ElevenLabs' own figures. The post shows a comparison chart against proprietary models but does not say in its text which tasks were measured or which alternatives the cost comparison uses, so they read as vendor claims rather than independent measurements.

The model line-up has since moved on. ElevenLabs' current model documentation lists newer models under ElevenLabs as the provider, such as GLM 5.2 and Qwen3.6-35B-A3B, and no longer lists GLM 4.5 Air or Qwen3-30b-a3b. The same page explains what the hosting arrangement means for data: models marked Hosted by ElevenLabs run on ElevenLabs-managed infrastructure, currently in the United States or the region of an enterprise deployment, and the original model developer does not receive the inputs or outputs. That arrangement, not the two models named at launch, is what still describes the offering.