贾扬清离职英伟达创立Intent Lab，推出性能提升6.3倍的AI推理系统

贾扬清正式宣布离职英伟达并创立新公司Intent Lab 1。该公司推出的首批AI基础设施产品包括Fleet系统、数据库引擎和分布式文件系统AgentFS 1。

在推理引擎性能方面，Fleet系统将GLM-5.2的输出速度从102 tokens/s提升至647 tokens/s，实现了6.3倍的性能提升 1。该系统通过内核层优化、运行时优化、通信优化和投机解码优化四类方案达成这一目标 1。

Intent Lab的其他产品也展现了显著成效。开发的数据库引擎已通过600万条SQLite兼容性测试 1，分布式文件系统AgentFS在部分指标上相比EFS快600倍，并进行了约190万个状态的形式化验证 1。贾扬清将Fleet定义为"全球第一个把意图变成生产级软件的自主团队" 1。

Jia Yanqing has officially established Intent Lab following his departure from NVIDIA, unveiling a suite of AI infrastructure technologies designed to automate software development at production scale 1. The company's flagship offering, the Fleet system, accelerates the GLM-5.2 inference engine by 6.3 times, boosting output speed from 102 tokens per second to 647 tokens per second 1. This performance improvement is achieved through four optimization approaches: kernel-level enhancements, runtime optimization, communication improvements, and speculative decoding 1.

Beyond the inference engine, Intent Lab has developed complementary infrastructure components to support autonomous AI systems 1. A SQLite-compatible database engine has passed 6 million test cases, while a distributed file system called AgentFS demonstrates performance improvements of up to 600 times compared to EFS in certain metrics and has undergone formal verification of approximately 1.9 million state transitions 1. Jia Yanqing characterizes Fleet as "the world's first autonomous team capable of converting intentions into production-grade software," positioning the platform as a comprehensive solution for end-to-end AI-driven software development 1.