Zyphra Technologies

Zyphra Technologies is an open-source and open-science artificial intelligence company headquartered in Palo Alto, California. Founded in 2020 by Krithik Puthalath and Danny Martinelli, the company focuses on developing multimodal AI models, learning algorithms, and systems. The team, which includes talent from organizations like Google DeepMind, Anthropic, and Nvidia, has expertise in AI models, learning algorithms, and systems infrastructure with a particular focus on inference efficiency and AI silicon performance.

The company develops autonomous agent platforms for enterprises, concentrating on conversational AI, automation, and personal assistants. Its core mission is to create AI models that are personalizable offline and can perform efficiently across a wide variety of hardware, from mobile phones and laptops to cloud systems and enterprise-grade GPUs from different vendors. Zyphra is building a multimodal general agent named Maia, which combines research in next-generation neural network architectures, long-term memory, reinforcement learning, and continual learning. Initially, Maia is targeting text and audio modalities.

As part of its commitment to open science, Zyphra has released several open-source projects. This includes Zamba, an SSM-hybrid foundation model designed to bring AI to more devices with lower inference costs. The company also released Zamba2-2.7B, a small language model (SLM) that offers increased speed and reduced memory overhead compared to other models. In June 2024, Zyphra introduced Zyda, a large-scale, open-source AI training dataset with 1.3 trillion tokens, designed to help researchers build large language models more efficiently by providing pre-filtered and deduplicated data. The company has received backing from institutional investors such as Future Ventures, Defined Capital, and Intel Capital.

Keywords: multimodal AI, artificial intelligence models, open-source AI, small language models, large language models, agent platform, conversational AI, offline personalization, inference efficiency, AI silicon performance, Zamba, Zyda, Maia, deep learning, machine learning, neural network architectures, reinforcement learning, continual learning, AI training data