MOSS-Live: Open-Source Full-Duplex Voice AI Is Here OpenMOSS released MOSS-Live on November 11, an open-source full-duplex voice AI that listens and speaks simultaneously, runs background tool calls without pausing conversation, and ships with a hosted API in beta and an Apache 2.0 GitHub repository. MOSS-Live scores 98.17% backchannel selectivity on the τ-Voice benchmark, 51.00% tool-use success and 95.68% tool selection F1 on FDB v3, and 4.88 out of 5.00 on post-interruption response quality, with turn-taking latency of 1.14 seconds normally and 1.38 seconds after an interruption. The release gives developers a self-hostable alternative to OpenAI's closed GPT-Live-1, which is faster at 0.798 seconds turn-taking latency and costs $0.05 per minute, though MOSS-Live still trails GPT-Live-1 on multi-turn task completion and Doubao on audio understanding. OpenMOSS released MOSS-Live on November 11 — an open-source, full-duplex voice AI that listens and speaks at the same time, runs background tool calls without pausing the conversation, and distinguishes a real interruption from a user just saying “mm-hmm.” It ships with a hosted API in beta, an Apache 2.0 GitHub repo, and benchmarks that hold up under scrutiny. For developers building voice agents, this is the first capable open alternative to GPT-Live-1 and Google’s Gemini Live that you can actually self-host. The Problem: Voice AI Is Still Half a Walkie-Talkie Most voice AI — including the majority of production voice agents in 2026 — works by stitching together a voice activity detector, a transcription model, an LLM, and a TTS engine. Each step adds latency. Total pipeline delay typically lands between 1.2 and 2.0 seconds, well above the 200–500ms window where human conversation feels natural. More importantly, users cannot interrupt. Any word spoken while the AI is talking is ignored. The agent waits for silence, then responds — walkie-talkie style. GPT-Live-1 https://platform.openai.com/docs/guides/realtime broke this pattern in September by processing audio natively, cutting turn-taking latency to 0.798 seconds. It works. But it is a closed API at $0.05 per minute, which becomes significant at scale, and it rules out any deployment where data cannot leave your infrastructure. MOSS-Live takes the same architectural bet — go full-duplex or stay awkward — and publishes the weights and the code. How MOSS-Live Solves It: Two Models, One Conversation The core of MOSS-Live is a dual-timescale architecture. A foreground model handles real-time listening, turn-taking decisions, and speech synthesis — it is optimized for latency and runs continuously. A background model handles reasoning, tool calls, and anything that takes longer than a moment. The two share memory and task state through a coordination layer, so the conversation continues while the background model works. This matters because it separates two fundamentally different requirements. Real-time speech needs to be fast. Reasoning does not need to be fast — it needs to be accurate, and the user does not notice the latency if the conversation keeps moving. By decoupling these, MOSS-Live avoids the tradeoff that breaks cascaded pipelines: you no longer have to choose between “respond quickly” and “use tools.” What It Actually Does Four capabilities define MOSS-Live in practice: - Interruption handling — MOSS-Live tracks whether you are genuinely interrupting or just producing backchannels mm-hmm, right, yeah . On the τ-Voice benchmark it hits 98.17% backchannel selectivity, meaning it almost never stops talking when you murmur acknowledgment. When you do interrupt, it stops and responds within 1.38 seconds. - Background tool use — Tools run asynchronously. The model keeps listening and responding while a tool call executes, then integrates the result when it comes back. On FDB v3, MOSS-Live achieves 51.00% tool-use success and 95.68% tool selection F1, tied for first in the evaluation. - Live video input — MOSS-Live accepts camera or screen-share input alongside audio, letting you build agents that understand what the user is looking at in real time. - Speech tracking — The model knows where it is in its spoken response. After an interruption, it picks up coherently rather than starting over or falling silent. Benchmarks: Honest Numbers MOSS-Live leads all six full-duplex metrics on the FDB evaluation suite. Post-interruption response quality scores 4.88 out of 5.00, the highest among evaluated models. Turn-taking latency is 1.14 seconds for a normal exchange and 1.38 seconds after an interruption. GPT-Live-1 is faster at 0.798 seconds for turn-taking — that gap is real and worth acknowledging. MOSS-Live also trails GPT-Live-1 on multi-turn task completion and Doubao https://www.doubao.com/en/ on audio understanding. This is not the top model on every metric. It is the top open-source model, with published evidence to support the claim. Developer Access Two paths are available today. The hosted beta API connects via WebSocket at wss://api.mosi.cn/v1/realtime with model ID moss-live-1.0 . Send 16 kHz mono PCM S16LE audio as continuous binary frames; keep your API key server-side only. A free tier is active during beta at platform.moss.cn https://platform.moss.cn/en/docs/scenarios/moss-live/ . For self-hosting, the full source is on GitHub under Apache 2.0 https://github.com/OpenMOSS/MOSS-Live . The community has already built a DeepSeek Harness plugin https://dshpluginhub.ai/en/plugins/dsh-moss-live that wires MOSS-Live as a voice front-end into the 276K-star DSH agent harness — meaning if you are already using DSH, adding voice is a plugin install. When to Use MOSS-Live vs. GPT-Live-1 GPT-Live-1 wins on raw latency and simplicity of integration. Use it if sub-second response time is a hard requirement and your data policy allows a closed API. MOSS-Live is the right choice if you need to self-host data compliance in healthcare, legal, or finance , if per-minute API costs are a concern at your call volume, or if you want to customize the model itself. The dual-timescale architecture also makes MOSS-Live a better base for research into voice agent behavior — you can inspect and modify both models independently. The full technical breakdown is in the Hugging Face blog post https://huggingface.co/blog/OpenMOSS-Team/moss-live and the official benchmark comparison https://openmoss.ai/blog/en/moss-live/ .