Most voice assistants still work like walkie-talkies: you speak, you stop, they answer. Boson AI wants to retire that rhythm. On November 5, 2026, the company introduced Higgs Realtime Duplex, which it describes as its first full-duplex interactive “omni” model. According to Boson’s announcement, it listens, watches and talks at the same time, shows a face of its own, and keeps thinking and calling tools in the background without pausing the conversation.
That is a meaningful shift for anyone building customer service agents, tutors or hands-free work assistants. Here is how Boson says it works, how it compares, and what to keep in mind.
What “full duplex” actually means
In a full-duplex system, input and output flow at once, the way they do in a phone call between two people. Boson notes that in human conversation, a reply typically lands within about 200 milliseconds of the other person finishing, and that people constantly decide when to speak: saying “mm-hm,” cutting in, helping when someone is stuck, and staying quiet when a remark is not meant for them.
According to the company, Higgs Realtime Duplex is designed to do the same. It can acknowledge you mid-sentence, answer before you finish when the timing is right, stop when you interrupt, and ignore background noise or side conversations. It also watches your camera or shared screen and responds with a face rendered in real time, in sync with its voice.
How it works: slicing time, not turns
The core design idea, Boson explains, is to cut every stream into 160-millisecond chunks. Your voice and video, tool results and replies from a larger model all arrive as chunks, and at each step the model reads the latest input chunk and writes the next output chunk: its voice, its face and its background work. Even silence is represented as chunks, so the model never stops listening.
That structure lets the model do several things at once:
- Think while you talk. It starts reasoning on early phrases instead of waiting for the full question, and can keep reasoning while it speaks.
- Call tools asynchronously. Search, knowledge bases, booking systems or CRM lookups run in the background, and the model can begin answering from partial results.
- Delegate. For harder tasks, the 45-billion-parameter “front” model hands work to a much larger backend model with full context, keeping the conversation going until the answer returns.
Boson contrasts this with Thinking Machines’ interaction models, where the front model leaves thinking and tool calls to a background model. Boson argues that letting its front model handle most requests itself avoids the long waits and unpredictable costs of calling a big reasoning model on every turn.
Built to be cheap enough to talk all day
Cost is an explicit design goal. Boson says its target is for an hour with the model to cost no more than one-tenth of an hour of a person’s time. To get there, the team squeezed each stream: one audio token per chunk in, two out, a camera frame only every six chunks, and two latents for the face.
According to the announcement, the audio and visual streams use about 2,600 tokens per minute, or roughly 2,600 to 5,000 with thinking and tool calls included. In production the model typically keeps the last five minutes of audio and video plus compacted text, about 13,000 to 25,000 tokens, with older history stored as text.
Voices and faces are swappable without retraining. A short reference audio clip sets the voice and a single image sets the face, fed into separate audio and video decoders rather than baked into the model.
Training and benchmark results
Boson says the front model is a 45B mixture-of-experts model, continued from a text-pretrained base on about 2 trillion tokens of text, audio and video, using its own NVIDIA H100 cluster in a data center the company built. Post-training combines specialist “teacher” models trained with reinforcement learning for conversation, interactive behavior, delegation and task completion, distilled into a single student. A self-improvement loop runs the agent in simulated conversations full of interruptions and background noise, then turns failures into new training data.
In Boson’s published comparisons against the Thinking Machines interaction model and OpenAI and Google realtime models, Higgs Realtime Duplex posted the top score on Big Bench Audio (98.9) and second-highest results on Audio MultiChallenge (44.9), τ-Voice task completion (69.8%) and IFEval (94.1). It trailed some rivals on Full-Duplex-Bench v1.5 (75.5 versus 80.1 for GPT-Live-1). Boson points out that its model is about six times smaller than the 276B-parameter TML Interaction Small, and reads the close results as a sign that data and training methods matter more than raw scale. These are company-reported figures; Boson says it reproduced more than half of the competitor scores to check its evaluation pipeline.
What it means
- More natural voice agents. Back-channel cues, interruptions and knowing when to stay quiet are what make a conversation feel human. Boson’s demos, including a drive-thru scene where the model stays silent while the driver orders, show why that matters.
- Assistants that work while they talk. Running tools and delegating in the background removes the dead air that makes today’s agents feel slow.
- Hands-free professional use. Boson’s engineering demo has the model comparing design revisions and coordinating a coding agent by voice alone, hinting at uses in workshops, labs and field work.
- Not available yet. The Duplex API is “coming soon” via a waitlist. Boson’s existing Higgs Realtime model is available today; its documentation describes a full-duplex voice API over WebSocket with turn detection, interruptions, tool calling and an OpenAI-compatible event protocol.
The bottom line
Higgs Realtime Duplex is a bet that talking will become the main way people work with AI agents, and that the winners will be models that feel natural, stay smart and stay cheap over hours of use. The architecture is thoughtful and the early numbers are competitive. The real proof will come when developers get API access and put it in front of actual customers.