Voice and NPCs
Higgs Realtime Duplex: a voice model with its own live face
Boson AI has introduced Higgs Realtime Duplex, a 45B model that listens, watches and talks at once in 160 ms steps, renders its own face, and keeps working on tasks in the background.
On 5 November, Boson AI introduced Higgs Realtime Duplex, which it calls its first full-duplex interactive omni model. It listens to your voice, watches your camera or shared screen, speaks, and renders a face of its own, all at the same time.
Boson says the model decides for itself when to speak, says “mm-hm”, answers before you finish, and stops when you cut in. It also recognises when a remark is not meant for it.
How it works
Boson cuts every stream into 160 ms chunks. Each step, the model reads the chunk that has just arrived and writes the chunk that plays next. Silence is represented as chunks too, so it never stops listening.
- Front model: a 45B mixture-of-experts model, trained on about 2 trillion tokens of text, audio and video.
- Background work: it can think, call tools and hand hard questions to a bigger backend model while the conversation continues.
- Voice and face: a short reference audio clip sets the voice and one reference image sets the face, with no fine-tuning. The face comes from a causal diffusion transformer (DiT) decoder, and Boson says it stays visually consistent for as long as the conversation runs.
- Cost: audio and visual streams use about 2.6K tokens per minute, or roughly 2.6-5K with thinking and tool calls. Boson’s design target is an hour of use costing no more than a tenth of an hour of a person’s time.
The numbers
Boson compared the model with Thinking Machines’ interaction model and with OpenAI and Google’s realtime and live models. These are Boson’s own figures, and it says it reproduced more than half of the other scores to check its pipeline.
- Big Bench Audio: 98.9, the highest in the table.
- IFEval: 94.1, second highest, behind one GPT configuration at 95.2.
- τ-Voice task completion: 69.8, second highest. GPT-Live-1 scores 74.5.
- Full-Duplex-Bench v1.5: 75.5, behind the Thinking Machines model (77.8) and GPT-Live-1 (80.1).
The results trade blows rather than sweep the table. Boson notes its front model is about a sixth the size of the 276B Thinking Machines model.
Access and limits
The Duplex API is “coming soon”, with a waitlist. Today, Boson’s end-to-end model with avatar output is available as Higgs Realtime, with pricing on its site. The post does not give a release date for the Duplex API, a price, or open weights. It also reports no voice-cloning or avatar-cloning results, because the other models do not offer those features.
Our take
For anyone building interactive worlds, the interesting part is not the benchmark table but the behaviour. A character that interrupts, backs off, ignores a side conversation and keeps its own face consistent is much closer to a person in the room than a character that waits for a turn.
It also shows the pattern we expect to spread: a small, fast front model that stays in the conversation, and a larger model working behind it. That is how a talking NPC can think about its plans without going silent. We explain the design problem in the end of the dialogue tree, and a similar approach appears in our post on MOSS-Live.