Blog Today, we are excited to announce Avartha. We are building the infrastructure to bring frontier-grade intelligence to real-time applications with open models that you control.
Siri launched in 2011 with the promise of helping us “get things done just by asking.” We want AI that can actually deliver on that promise, not just sound good. We now have models that can write software and reason through complex problems, but developers building voice agents have had to trade that capability for responsiveness. Even in mid-2026, GPT-4.1 remains the most popular model for building voice AI applications, simply because it is the only widely available model which has sub-1 second latency [5].
We believe that voice is poised to be the dominant interaction surface for all of computing in the future, and we want to build voice agents that operate at the frontier of AI capabilities and are a worthy complement to the intellect of their users. As the first step towards this goal, we are releasing Ava Zero<sup>1</sup>. Ava is built entirely on open models and ranks #1 in task completion on Sierra’s τ³-Voice benchmark. τ³-Voice tests whether an agent can reason accurately and execute long, multi-step trajectories under real-world conditions. Ava completes 86.0% of tasks across six languages, ahead of OpenAI’s GPT Live 1 at 79.9%, xAI’s Grok at 71.4%, and Google’s Gemini at 68.5%, with the highest generation quality. Open models are now at the frontier of voice AI.
| Rank | Model | Task completion | Interaction | Generation quality |
|---|---|---|---|---|
| #1 | Ava ZeroAvartha | 86.0% | 25.0% | 62.2% |
| #2 | phonic-1Phonic | 83.2% | 36.2% | 31.3% |
| #3 | gpt-live-1OpenAI | 79.9% | 32.1% | 61.6% |
| #4 | grok-voice-think-fast-2.0xAI | 71.4% | 52.6% | 17.7% |
| #5 | gemini-3.8-live-extended-thinkingGoogle | 68.5% | 15.8% | 37.8% |
[τ-bench leaderboard](https://taubench.com/leaderboard/?benchmark=voice), Multilingual, read October 29, 2026.
Reimagining Infra for Real-Time AI #
In traditional AI applications like coding and chatbots, latency affects the user experience but not the result; a slow answer is still a correct one. In real-time applications, meeting the deadline is part of completing the task. The capabilities of the application are therefore bounded by the models that fit within its latency budget. We thus believe that building real-time agents that are both correct and fast is fundamentally an infrastructure challenge, requiring ultra-low-latency inference and high tail reliability.
Today, agent infrastructure is fragmented across agent harness platforms and inference providers. Both these layers are built independently by different teams and stitched together with narrow interfaces. Such interface boundaries come at a cost to both performance and efficiency, as shown by past research on databases [1] and operating systems [2–4]. In contrast, we have co-designed the entire real-time agent infrastructure, from the harness to the kernels, as one system optimized for end-to-end task performance: a full-stack agent factory.
We built our inference engine from scratch using first principles to deliver the best performance and efficiency for voice AI applications. Since we deeply understand the behaviour of our application, we were able to drop generalized assumptions and design every component of the system in a manner that exploits the specific characteristics of the workload. We can support up to an order of magnitude more concurrent conversations per GPU than vLLM and SGLang across text-to-speech, speech-to-text, and language models under strict latency constraints. This allows us to leverage frontier-grade models like GLM-5.3 to drive the voice agent without compromising on latency, while the rest of the industry still operates with models an order of magnitude smaller.
As a direct consequence of this, we were able to completely reimagine how a voice agent harness operates and construct it in a manner that can fully utilize the intelligence at our disposal. Our τ³-Voice leaderboard submission for Ava uses a combination of Qwen3-ASR for transcription, Fish Audio Drama 3 for speech generation, GLM-5.3 as the anchor, and DeepSeek 4.1 Flash for backend workers.
We are also able to translate the efficiency of our stack into better pricing for our consumers. Ava costs just $2.50 per hour of conversation, one-half to one-third the cost of comparable closed-source alternatives.
The Avartha team #
Avartha spun out of the Systems for AI Lab at Georgia Tech. Our team introduced some of the most fundamental techniques that helped shape the modern inference stack. Between us, we have co-created chunked prefills (the default scheduling algorithm for all major inference systems), Ray, KV-cache parallelism, and chunked pipeline parallelism. We also built the most popular inference simulator and pioneered a GPU-free inference emulation technique to enable fast and accurate deployment optimization and agentic development loops.
It is our belief that “One Size Fits All” is an idea whose time has come and gone. The next generation of inference infrastructure is going to be hyperspecialized by workload, and we are intent on building the best systems for real-time applications.
We’re just getting started #
Our mission is to make real-time AI more capable, faster, and more reliable. We are working hard to push the boundaries of voice systems and solve open problems, including multi-speaker dialogue and personalization. Zooming out, we believe that real-time intelligence systems will reshape our physical world, and we are committed to building frontier-grade infrastructure to keep open models at the forefront across modalities.
We’re hiring exceptional builders who want to work on hard systems and machine learning problems and put their ideas into the hands of developers. Explore our open roles.