Boson AI unveils Higgs Realtime Duplex, a 45B model that listens, watches and talks at once

Boson AI, the Santa Clara company led by co-founder Alex Smola, introduced Higgs Realtime Duplex on November 5. In a post on its blog titled "Face to Face, Working in the Background", the company presents it as the successor to its Higgs Realtime speech model and as its first full-duplex interactive omni model: a system that keeps taking in audio, video and text while it speaks, reasons and animates a face of its own.

What the model does

According to Boson AI's description, as relayed by RuntimeWire and TLDR, the model is a 45-billion-parameter mixture-of-experts network that the company calls its "front model". It consumes voice and video in 160-millisecond chunks, which is meant to let it respond and be interrupted with little delay instead of waiting for a user to finish a turn. The company says it can call tools asynchronously and pass heavier work to a larger backend model while the conversation continues; that backend model is not named in the material reviewed. The company also says the model renders a live face while it talks, but the pages reviewed give no technical detail on how this works.

Boson AI also frames the release around cost. RuntimeWire reports the company's goal that an hour of the 45B front model should cost no more than a tenth of an hour of human time. That is a stated target, not a measured figure.

What is still unclear

Several basic points could not be verified. Boson AI's full blog post was not accessible for this article, and the secondary pages consulted give no license, no indication of downloadable weights and no benchmark results; a Papers with Code entry for the post mentions an instruction-following benchmark but lists no scores. The number of active parameters per token is not stated. AI Weekly reports that API access is coming through a waitlist, while the earlier turn-based Higgs Realtime remains available. No pricing has been published.

The launch extends a line of voice products from the company. In recent weeks Boson AI's blog has announced live streaming of its Higgs Avatar into voice-agent calls and a guide to building a low-cost conversational device with Higgs Realtime. Full-duplex systems that listen and speak simultaneously are an active research area, and independent tests will be needed to judge how Boson AI's latency and cost claims compare with other real-time voice models.