Boson AI Higgs Realtime Duplex: 45B Voice-Video Model

Boson AI officially launched Higgs Realtime Duplex on November 5, 2026, marking a significant step forward in conversational AI architecture. The 45B-parameter mixture-of-experts model processes voice and video in 160 millisecond chunks while simultaneously speaking, reasoning, calling tools, and delegating complex tasks to larger backend models. Trained on approximately 2 trillion tokens, the model scored 98.9% on Big Bench Audio and 94.1% on IFEval benchmarks, with API access currently available through a waitlist.

Full-Duplex Architecture and Processing

Unlike turn-based systems, Higgs Realtime Duplex maintains continuous conversation flow by processing audio and video in overlapping chunks. The model encodes each 160 ms segment into conditioning vectors for speech generation and uses a causal diffusion decoder for video output. This architecture enables the model to respond within the 200 millisecond window typical of human conversation, matching natural interaction patterns including interruptions, acknowledgments, and contextual silence.

The front model handles ongoing reasoning threads that persist across chunks, building on partial information as each phrase arrives rather than waiting for complete questions. When visual context is needed, a causal DiT diffusion decoder generates video frames from reference images and the two most recent latents, maintaining facial consistency throughout extended conversations.

Training and Distillation Methodology

Boson AI trained the model on its own NVIDIA H100 cluster using a stack based partly on Megatron, continuing from a text pre-trained base through roughly 2 trillion tokens of multimodal data. The post-training phase employs multi-teacher on-policy distillation, where specialist teacher models trained via reinforcement learning on grouped tasks (natural conversation, interactive behavior, dynamic delegation, task completion) are distilled into a unified student model.

The training incorporates recursive self-improvement loops where the agent system interacts with simulated environments, analyzes failures, synthesizes targeted training data, and retrains. These simulations reproduce realistic interaction patterns including overlapping speech, background noise, and interruptions to teach natural conversational behavior at scale.

Asynchronous Tool Calling and Delegation

The model initiates tool calls for search, knowledge bases, CRM lookups, and custom functions as early as possible, processing partial results while maintaining conversation flow. When tasks exceed the front model's capabilities, it delegates to a larger backend model (optimized for frontier models) while continuing to handle other conversation threads. This asynchronous architecture allows the 45B front model to balance cost efficiency with access to more capable reasoning when needed.

Boson AI designed the system to target a cost-per-hour ratio of one-tenth of human labor time, positioning it for extended daily agent interactions. The mixture-of-experts architecture at 45B parameters represents a deliberate balance: larger models score higher on benchmarks but increase operational costs, while the ability to delegate complex reasoning to backend models maintains capability on difficult tasks.

Release Date and Availability

Higgs Realtime Duplex was officially released on November 5, 2026, with API access opening through a waitlist system. The previous turn-based Higgs Realtime model remains available for applications that do not require full-duplex interaction. Boson AI built its own data center infrastructure to support training and deployment, giving the company direct control over the compute stack powering both training runs and inference serving.

What This Means

Higgs Realtime Duplex demonstrates a production-ready approach to full-duplex conversational AI that combines multimodal processing, continuous reasoning, and cost-efficient architecture. By training specialist teachers and distilling their capabilities into a single model, Boson AI has created a system that handles the overlapping demands of listening, speaking, thinking, and acting without artificial turn-taking boundaries. The 45B parameter size, asynchronous tool integration, and backend delegation strategy position the model as a practical foundation for extended agent interactions across enterprise and consumer applications.