Boson Higgs Duplex: 98.9 Audio, 75.5 Full-Duplex

Conversation without turns

Most voice assistants still organize conversations around turns: the user speaks, the system processes the request, and the assistant replies. Boson's Higgs Realtime Duplex is designed to process incoming audio, video and text while generating speech and facial output in parallel. Boson describes a 160-millisecond chunking cadence, including chunks containing silence. That is the model's streaming interval, not a published measurement of end-to-end conversational latency. The design is intended to support interruptions, backchannels and continuous visual awareness without requiring the user to finish speaking first.

Boson's demonstrations illustrate backchannel responses, interruptions, early answers and selective silence when speech appears directed at someone else. The examples show what the architecture is intended to support, but the announcement does not establish how reliably each behavior works across different speakers, environments and conversations. It also sets the voice and the face from references at deployment time: a short audio clip conditions the speech decoder, a single image conditions the face decoder, and neither requires fine-tuning.

The part that works in the background

The design choice worth attention is where the thinking sits. Boson's front model reasons on its own, in a thread that runs alongside listening and speaking. It fires tool calls as early as it can, starts answering from partial results, and hands only the tasks beyond it to a much larger backend model, with the full shared context, while it keeps the conversation going on other topics.

Boson contrasts its design with Thinking Machines' approach, which places more reasoning and tool execution in a separate background model. Boson argues that giving the 45B front model its own reasoning and tool-use capabilities can reduce the need for backend delegation. That is an architectural claim, not a demonstrated cost or latency advantage over Thinking Machines. The announcement does not publish a controlled comparison of backend invocation frequency, response latency or operating expense. Boson's 45B front model has approximately one-sixth the nominal parameter count of Thinking Machines' 276B Interaction Small. That comparison concerns reported model size, not active parameters per token, compute requirements or inference cost.

Boson estimates approximately 2,600 tokens per minute for the base audio-and-video streaming workload, rising to roughly 2,600–5,000 tokens per minute when reasoning, tool calls and their results are included. At those rates, five minutes of recent interaction would represent approximately 13,000–25,000 tokens. These are Boson's architectural workload estimates, not measured billing totals or a published context-window limit. The company also describes a design objective of keeping an hour of interaction below one-tenth the cost of an hour of human time. No Duplex API price or independently measured hourly operating cost is published, so that objective cannot yet be evaluated.

First on audio QA, behind on full-duplex

Boson's published comparison table gives Higgs Realtime Duplex a Big Bench Audio score of 98.9, above the other configurations listed, including Thinking Machines at 96.5 and Gemini 3.8 Live with extended thinking at 98.0. That makes Duplex the leader in Boson's reported table for this benchmark, not necessarily across every available model or evaluation setup.

On Full-Duplex-Bench v1.5, Boson reports 75.5 for Duplex, compared with 77.8 for Thinking Machines and 80.1 for GPT-Live-1. Duplex ranks behind both in the configurations shown. The table combines Boson's results with competitor scores drawn from published reports and leaderboards, some of which Boson says it reproduced. It should not be read as a single independently administered head-to-head test.

Boson's table reports 44.9 on Audio MultiChallenge, below a GPT Realtime-2 configuration at 48.5; 69.8% task completion on τ-Voice, below GPT-Live-1 at 74.5%; and 94.1 on IFEval, below a Realtime-2 configuration at 95.2. On streaming video QA, Duplex scores 56.0, compared with Realtime-2 configurations scoring 57.5 and 58.2. These figures show competitive performance across several tasks, but they do not establish consistent leadership. Differences in model configuration, thinking budget and evaluation provenance limit direct comparisons. None of this makes Duplex weak. It makes the launch claim narrower than the headline: a much smaller front model trading blows with larger and higher-thinking configurations, and leading the configurations shown on Big Bench Audio.

Two cautions belong with the table. The competitor scores come from published reports and leaderboards, with Boson saying it reproduced more than half of them to check its evaluation pipeline, which also means nearly half were not reproduced. And GPT and Gemini reach many of their best scores with thinking set high, a setting that costs time and tokens Boson's per-minute budget is designed to avoid. The comparison is informative. It is not a controlled race.

What you can actually use today

Nothing called Duplex. The API is a waitlist, the weights are not released, and there is no date in the announcement. What Boson sells today is Higgs Realtime, the previous model, which the announcement describes as an end-to-end model with avatar output at published rates. A buyer comparing the two should be careful: every benchmark and every behavior above belongs to the model that is not out yet.

The architecture and Boson's reported benchmarks are available to study, but developers cannot yet evaluate Duplex through a generally available API or released weights. The existing Higgs Realtime product is a separate, earlier offering. Whether Duplex delivers a better interactive experience, lower latency or lower operating costs than competing live models remains unproven until comparable hands-on testing and commercial pricing are available.

Sources

[1] Higgs Realtime Duplex: Face to Face, Working in the Background, Boson AI, 5 November 2026 (primary)Read source
[2] Talking while thinking, Alex Smola (Boson AI), 5 November 2026Read source
[3] Higgs Realtime Duplex coverage and benchmark summary, Emergent, 6 November 2026Read source

THE ALFW VERDICT

Higgs Realtime Duplex is interesting for its architecture: a 45B front model that processes continuous multimodal input, generates speech and facial output, and performs some reasoning and tool use without always handing the conversation to a larger backend. The benchmark story is more mixed. Boson reports the highest Big Bench Audio score among the configurations in its table, at 98.9, but its 75.5 on Full-Duplex-Bench v1.5 trails the listed Thinking Machines and GPT-Live-1 results. The comparison combines reported and partially reproduced competitor scores, so it is evidence of competitiveness rather than a controlled demonstration of superiority. The commercial claims remain untested. Boson's low-cost target is not a published price, the 160 ms chunking cadence is not an end-to-end latency result, and Duplex access is limited to an API waitlist. The architecture deserves attention now. Its conversational reliability, operating economics and competitive advantage still need independent evaluation.