OpenRouter团队在Hacker News上宣布推出开源模型网关OpenRouter，可统一管理自托管、前沿及开源模型 1。该网关支持所有主要推理提供商及1000多种模型，允许用户混合使用本地与市场模型，并可选择使用流量训练模型 1。作为开源项目，它支持自托管或使用托管版本，且托管服务不收取任何额外加价 1。
在性能与技术实现上，OpenRouter采用Rust原生构建并专为并发设计，实现了极低的延迟 1。数据显示，自带密钥（BYOK）请求增加的延迟低于1毫秒，而在使用Experiential提供的提供商密钥时延迟增加也低于2毫秒 1。为优化成本与质量，系统每日通过codex agent刷新模型并开启拉取请求 1。同时，该网关利用标准化OTel traces挖掘代表性真实任务，使用文本世界模型模拟rollout并结合LLM judge进行评估 1。最终，系统通过在提示词嵌入上拟合最近邻分类器，为每个请求选择最优模型 1。
The OpenRouter team has introduced an open-source model gateway named OpenRouter on Hacker News, designed to unify the management of self-hosted, frontier, and open-source models 1. Built natively in Rust and optimized for concurrency, the gateway ensures ultra-low latency, adding less than 1 ms of delay for Bring Your Own Key (BYOK) requests and under 2 ms when Experiential provides a provider key 1. It supports all major inference providers and over 1,000 models, which are refreshed daily via a codex agent that automatically opens pull requests 1.
To optimize both cost and quality, OpenRouter utilizes standardized OTel traces to extract representative real-world tasks 1. The system then simulates rollouts using a text world model combined with an LLM judge, and fits a nearest-neighbor classifier on prompt embeddings to select the optimal model for each request 1. The platform is fully open-source, allowing users to self-host or use the managed version with zero markup, and enables the mixing of local and market models 1. Additionally, users have the option to use their traffic to train and improve the models 1.
评论
还没有评论，欢迎留下第一条。