AI Briefing
KOSign in

Boson AI Releases Higgs Realtime Duplex: 45B MoE Model for Full-Duplex Voice and Video Interaction

·2026.11.07 05:23

Key point

The 45B parameter model processes audio, video, and text in 160ms chunks while delegating complex reasoning to a larger backend model to maintain low latency.

Details

  • Boson AI released Higgs Realtime Duplex, a 45B parameter Mixture-of-Experts (MoE) model that operates in full-duplex mode, listening, watching, and speaking simultaneously rather than in turn-based sequences.
  • The model processes inputs in 160ms chunks, generating approximately 2.6K tokens per minute for audio and video, with a context window of 13K to 25K tokens covering the last 5 minutes and compacted history.
  • Complex reasoning, tool calls, and delegation are handled in the background by a larger backend model, allowing the front model to start speaking before tool results are complete, unlike competitors where all reasoning goes to the background.
  • Benchmarks show Higgs Realtime Duplex trails OpenAI's best settings on five of seven metrics by 1 to 5 points, with the largest gap in turn-taking performance on Full-Duplex-Bench.
  • The system uses zero-shot voice and face cloning via reference clips and images, trained on 2T tokens of text, audio, and video using a Megatron-based stack on H100 clusters.
  • The design targets a cost where one hour of model time costs at most a tenth of an hour of human time, prioritizing data and training efficiency over raw parameter count.
moemodel-releasevoice-aifull-duplexreal-time-inferenceboson-aihiggs-realtime-duplex
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

PyTorchKRThinking Machines Lab Releases Active 17B Inkling 2 Small Tinker and Plans to Open WeightsRedditChina Telecom releases Xing4.0-29B-A4B, a 29B MoE model trained on Ascend NPUsRedditLinum-AI Releases P-JiT 21K Text-to-Image Model Trained for $21KElevenLabsVoice AI infrastructure: The 2026 stack and architectureHacker NewsThinking Machines AI introduces Inkling 2 Small, planning Apache 2.0 open weights release next weekTLDR AIMeta rolls out voice and expressive avatars in the Muse app