Distributed AI infrastructure for inference and reinforcement learning
Gradient is an AI infrastructure lab building three interconnected systems: Parallax (distributed inference across heterogeneous hardware), Echo (asynchronous reinforcement learning framework), and Lattica (peer-to-peer networking layer). The stack reveals hardware-agnostic design—PyTorch, JAX, TensorFlow alongside Rust and Go—optimized for deploying AI across mixed device pools rather than centralized data centers. Early traction on Lattica (millions of peers) and Echo-2's 90%+ cost reduction in post-training suggest the team is past conceptual stage.
Gradient operates as an AI research lab focused on decentralized AI infrastructure. The company's three core products address specific bottlenecks: Parallax handles automatic model sharding and request routing across GPUs and consumer hardware; Echo decouples RL training from rollout collection to reduce post-training compute; Lattica provides global peer-to-peer coordination. Founded in 2024 and backed by Pantera Capital, Multicoin Capital, and HSG, the company is currently hiring research and engineering talent from Singapore. Active projects span distributed inference systems, incentive mechanism design, and partnership development around the open infrastructure thesis.
Core languages: Python, Rust, Go. ML frameworks: PyTorch, JAX, TensorFlow. The mix reflects hardware-agnostic, distributed-systems thinking required for inference across heterogeneous devices.
Parallax is a distributed inference engine that automatically shards models and routes requests across mixed hardware (NVIDIA GPUs, Apple Silicon, other devices), enabling local, offline AI deployment.
Echo-2, Gradient's distributed RL framework, reduced post-training costs for a 30B parameter model by over 90% (from ~$4,490 to ~$425) by decoupling rollouts from training asynchronously.
Other companies in the same industry, closest in size