Avartha launches Ava Zero voice agent to compete with OpenAI and Google

Bridging the gap in voice AI

For a long time, developers building voice AI agents faced a difficult trade-off between intelligence and speed. While models like GPT-4.1 are popular because they can respond in under one second, they often lack the reasoning depth required for complex tasks. Avartha aims to solve this by creating voice agents that can actually deliver on the promise of getting things done through conversation.

The company recently launched Ava Zero, an open-model voice agent designed to operate at the frontier of AI capabilities. Unlike previous models that might sound human but struggle with multi-step instructions, Ava Zero is built to reason accurately. It is intended to be a worthy complement to human intellect rather than just a simple chatbot.

By using open models, Avartha provides developers with more control over their infrastructure. The goal is to move beyond simple voice commands toward agents that can write software and navigate real-world conditions. This shift marks a move toward making voice the dominant way people interact with computers.

Outperforming the tech giants

Avartha tested Ava Zero against several major industry leaders using the τ³-Voice benchmark. This specific test measures an agent's ability to reason and execute long, multi-step trajectories in real-world scenarios. Ava Zero completed 86.0% of the tasks across six different languages, setting a new standard for open-weight models.

The results show that Ava Zero outperformed several high-profile closed models from the biggest tech companies. As avartha.ai reported, Ava Zero beat OpenAI’s GPT Live 1, which had a 79.9% completion rate, and xAI’s Grok, which finished 71.4% of the tasks. It also surpassed Google’s Gemini, which completed 68.5% of the benchmark tasks.

In addition to its high completion rate, the model was noted for having the highest generation quality among the tested systems. This suggests that the open-weight approach is no longer a compromise on quality. It provides a viable alternative for companies that want to avoid the limitations of proprietary APIs while still achieving top-tier performance.

A custom full-stack factory

To achieve these results, Avartha moved away from the standard way of building AI applications. Most current systems stitch together different platforms for the agent's brain and the inference engine, which can create bottlenecks. Avartha instead co-designed the entire infrastructure as a single system, which they call a full-stack agent factory.

By building their own inference engine from scratch, the team could optimize for the specific needs of voice AI. This allowed them to drop generalized assumptions and focus on ultra-low-latency inference and high tail reliability. Their system can support up to an order of magnitude more concurrent conversations per GPU than common frameworks like vLLM and SGLang.

This specialized infrastructure allows them to use more powerful models like GLM-5.3 to drive the agent. Because the system is so efficient, they do not have to shrink the model size to meet latency requirements. This efficiency also translates to a lower price point of $2.50 per hour of conversation for the end user.

The team behind the tech

The Avartha team originated from the Systems for AI Lab at Georgia Tech. The researchers involved have contributed to some of the most fundamental techniques in the modern inference stack. Their work includes the development of chunked prefills, which is the default scheduling algorithm for all major inference systems, as well as Ray and KV-cache parallelism.

The team also pioneered a GPU-free inference emulation technique to help with fast deployment and development loops. They built the most popular inference simulator to help developers optimize their agents before moving to production. This deep background in systems engineering is what allows them to build specialized hardware-level optimizations for real-time tasks.

Looking forward, the company plans to tackle more complex problems like multi-speaker dialogue and personalization. They believe that the next generation of inference will be hyperspecialized by workload. Their mission is to ensure that open models remain at the forefront of the AI field as real-time intelligence begins to reshape the physical world.