Is Higgs Realtime Duplex state of the art?Pending benchmarks.Announced Nov 5, 2026; no independent results yet.
Boson AI's research preview of the next generation of Higgs Realtime: a full-duplex interactive omni model that listens, watches and talks at the same time while rendering a live face, and that can think, call tools and delegate to a larger backend model without pausing the conversation. Front model is a 45B MoE trained on about 2 trillion tokens of text, audio and video, running in 160 ms chunks. An API for Duplex is on a waitlist; the current Higgs Realtime speech-to-speech API remains available separately.
The details so far
- Lab
- Boson AI
- Announced
- Nov 5, 2026
- API
- Waitlist
- Open weights
- Closed · None stated for Duplex
- Parameters
- 45B MoE (front model)
- Role
- Full-duplex interactive omni model: continuous audio and video in, speech and a rendered face out, with background thinking and tool use
- Chunking
- Streams cut into 160 ms chunks for input (voice, camera) and output (voice, face, background work)
- Architecture
- 45B mixture-of-experts front model; can hand hard turns to a larger backend model with shared context while keeping the conversation going
- Training
- Continued training on about 2 trillion tokens of text, audio and visual data; multi-teacher on-policy distillation and a recursive self-improvement loop (per Boson)
- Face and voice
- Zero-shot: a short reference audio clip sets the voice and a reference image sets the face at deployment time (not baked into the weights)
- Access
- Research preview / demo; Duplex API on waitlist ("coming soon" per the Nov 5 blog). Current Higgs Realtime remains the hosted speech-to-speech API
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: Big Bench Audio accuracy 98.9; Audio MultiChallenge APR 44.9; Full-Duplex-Bench v1.5 average 75.5 (Boson's published comparison table).
Lab claim: QIVD streaming accuracy 56.0; ProactiveVideoQA PAUC@ω=0.5 of 31.9; τ-Voice task completion pass@1 69.8%; IFEval accuracy 94.1.
Lab claim: design target is that an hour with the model costs at most one-tenth as much as an hour of a person's time.
Sources
LABHiggs Realtime Duplex blog (Boson AI)boson.ai/blog/higgs-realtime-duplex · Nov 5, 2026LABBoson AI on X: Higgs Realtime Duplexx.com/boson_ai/status/2118403814403580173 · Nov 5, 2026Tracking ends Feb 3, 2027 unless the API or weights become available.