Boson AI unveils Higgs Realtime Duplex, a 45B full-duplex voice-video model

Boson AI positions Higgs Realtime Duplex as a front model that handles most conversational work itself, reserving a larger backend model for harder turns. The company argues that delegating all reasoning and tool use to a bigger model is simpler to train, but a large reasoning model may take ten seconds or more to respond, making per-turn cost hard to control. Instead, the 45B front model thinks, calls tools, and delegates only when needed, while the backend works in the background.

Training relied on multi-teacher on-policy distillation and a recursive self-improvement loop that has the agent system interact with simulated environments, review failures, and synthesize targeted data. Boson also cut token usage deliberately: one audio token per chunk in, two out, a camera frame every six chunks, and two latents for the face. That design keeps the streams at about 2.6K tokens per minute, rising to roughly 2.6–5K with thinking and tool calls, and roughly 13K–25K for the recent five-minute window in production.

Benchmarks against other realtime and live models show mixed results, which Boson reads as a sign the field is still early. Notably, its 45B front model is compared with a 276B model from Thinking Machines, about six times larger, and the results trade blows. That framing suggests data, architecture, and training methods may matter more than raw scale for interactive voice-video agents.