Boson AI announces Higgs Realtime Duplex, a 45B MoE full-duplex voice model
Boson AI says Higgs Realtime Duplex is a 45B mixture-of-experts model that listens, watches and talks simultaneously while thinking, calling tools and delegating harder tasks to a larger backend model in the background. The vendor describes a 160 ms chunking scheme covering audio in and out, periodic video frames and face-output latents, plus multi-teacher on-policy distillation and a self-improvement loop in simulated environments. The release is not open weights, the Duplex API is waitlisted while the earlier turn-based Higgs Realtime remains available, and Boson AI concedes OpenAI's best settings lead on five of seven benchmarks, with Full-Duplex-Bench where it trails most.
The post offers a detailed first-party account of running audio, video, text and tool calls in one continuous stream with background delegation, a design contrast to turn-based voice models and to Thinking Machines' background-model approach.
Disclosure: I'm at Boson AI and this is our release. Not open weights; the Duplex API has a waitlist, and the previous turn-based Higgs Realtime is available now.
Most voice models are still turn-based: listen, wait for end of turn, think, speak. Duplex runs all streams at once and keeps working in the background while it talks.
Streams. Everything is cut into 160 ms chunks:
- Audio in: 1 token per chunk (6.25 tokens/s)
- Audio out: 2 tokens per chunk (12.5 tokens/s)
- Video in: one ~12-token camera frame every 6 chunks
- Face out: 2 latents per chunk
- Text in and out alongside
Silence is represented as chunks too, which is what the model uses to decide when to backchannel, interrupt or stay quiet. That comes to ~2.6K tokens per minute for audio and video, 2.6K to 5K with thinking and tool traces (progressively compacted). The last 5 minutes plus compacted text history is 13K to 25K tokens; older history is kept as text.
#Boson AI#Higgs Realtime Duplex#Higgs Realtime#TML Interaction Small#Thinking Machines#OpenAI#Google