Developers introduced the non-autoregressive model von-1.0

A team of researchers has released von-1.0 — an open decision-making system built on a non-autoregressive architecture. The model is focused on high-speed data processing and real-time classification tasks.

During testing, von-1.0 demonstrated an accuracy of 93.5% on eight different benchmarks. The key feature of the system is its speed: the response time is less than 25 ms. Such performance allows the solution to be used in tasks that require minimal delay in processing incoming signals.

According to the authors, the model outperforms existing analogs, including GLiNER2 and TypeSafe Jev, in terms of efficiency. The architectural approach that excludes autoregression allows the system to process information in parallel, which significantly reduces computational costs compared to traditional transformers.

The model release is available on the Hugging Face platform under an open license. The developers note that von-1.0 is optimized for integration into systems where the speed of decision-making is critical, rather than the generation of long text sequences.

The emergence of such specialized models indicates a trend towards transitioning from universal chatbots to specialized tools for automating decision-making. The use of non-autoregressive methods is becoming a standard for tasks that require predictable response times and high classification accuracy.