Skip to navigation
Models

Models overview

Compare Think-Fast and phoneLLM mini, the Deepgram LLMs available in Early Access.

Early Access. Deepgram LLMs are available only to current Deepgram contract customers enrolled in the Early Access program (talk to sales for access). Usage is free during Early Access, requests run on the global endpoint only, SDKs do not support the endpoint yet, and the API can change before general availability. Do not run production traffic on it.

Both models use the same endpoint and the same request and response format. To switch models, change the model field.

Model comparison

Think-FastphoneLLM mini
model valuethink-fastphonellm-mini-alpha-2-fp8
Use it forThe only LLM in a cascaded voice agentVoice agent pipelines where speed and size come first
StreamingYesYes
Function callingYesYes
ReasoningOff by default. Set reasoning_effort to none or omit it.Off by default. Set reasoning_effort to none or omit it; other values put reasoning text in the reply.
Model Improvement ProgramAlways included. mip_opt_out: true is rejected.Included by default. mip_opt_out: true is supported.
Voice Agent think providerdeepgramdeepgram

Think-Fast

Think-Fast is an instruction-following model built for voice conversations. It follows a system prompt and calls your function tools, so it can be the only LLM in a cascaded voice agent. It’s both fast and capable, and it writes replies meant to be spoken by a TTS provider.

It doesn’t reason by default. Leave reasoning_effort unset or set it to none: other values add reasoning tokens and latency before the reply starts. Put the steps you want it to follow directly in the system prompt. See Prompting Deepgram LLMs.

phoneLLM mini

phoneLLM mini is an open-weights text model from Pipecat, trained for voice agent pipelines. Choose it when speed and size are the top priority. It is a 30B-parameter mixture-of-experts model with 3.5B active parameters, based on NVIDIA Nemotron 3 Nano.

  • Tool calling: supported. The model was trained on synthetic voice agent sessions, including Pipecat Flows sessions.
  • Reasoning: off. Set reasoning_effort to none or omit it; other values put reasoning text in the reply, ending in </think>, which a voice app would read aloud. In the Voice Agent, the same setting is agent.think.provider.reasoning_mode.
  • Temperature: Pipecat recommends 0.0.

For model details and benchmarks, see Pipecat’s phoneLLM Alpha 2 announcement.

Availability

During Early Access, both models run on the global endpoint:

https://agent.deepgram.com/v1/think

Regional endpoints are coming soon.