Deepgram introduced Think-Live LLM and Think-Fast LLM on October 29, assigning conversation timing and task execution to separate models. Its launch account says Think-Live is generally available in the Voice Agent API, while Think-Fast is in early access on the Deepgram LLM API. Those are different access stages.
One model follows the conversation; another does the work
Deepgram describes Think-Live as deciding when an agent should listen, wait or respond. Think-Fast handles instructions, tool calls and replies intended to be spoken. The company says developers can pair either with other models rather than adopting both. These are provider-described capabilities; we have not measured interruption quality or response latency.
The same post says the broader Voice Agent Platform, built around Pipecat, is still forthcoming, with timing to follow. Its announcement should not be confused with general availability of every component. The model access described here is through APIs; the announcement is not a downloadable-weight release.
Daily is joining Deepgram; Pipecat’s openness is a separate question
Deepgram also announced that it is acquiring Daily, combining speech models with Daily’s network infrastructure and Pipecat orchestration. We are reporting an acquisition announcement, without independently verifying legal completion.
Daily’s own statement says its products will continue to be sold, supported and developed. The companies say they are establishing an independent Pipecat Foundation, with a future steering committee representing the wider ecosystem, including Deepgram competitors. That is a governance commitment to watch, not an independent audit of a functioning foundation.
The current Pipecat framework license is BSD-2-Clause. Its public repository documents interchangeable model and transport integrations. That framework license does not make every connected model or hosted service open source or free.
The context standard is already documented—and explicitly alpha
RCS 0.1 documentation describes inline context for how a user sounded and how a reply should be delivered. It warns of potentially incompatible changes before 1.0. Speaker identification and conversational intent remain on its roadmap; the specification lists implementation in Pipecat 1.15.0 and later.
The linked demonstration repository uses several providers and requires API credentials. It is an integration example, not evidence of entirely offline operation or a zero-latency conversation. We inspected the sources without running the demo.
For builders, the useful development is the separation of conversation control, task work and expressive context. Whether that improves a real application still needs testing against its users, network, tools and chosen models.
Original reporting from October 29 primary announcements, repository license and protocol documentation. Availability is attributed to Deepgram. No API call, audio evaluation or model installation was performed.
