Evo 在 B300 上优化 V4.1 Flash 推理：吞吐 +5%、P90 交互 +15%
Evo 从 vLLM 初始 V4.1 Flash 支持出发，在 AgentX 评测中输出吞吐 189.37 vs 181.85 tok/s、P90 交互 664 vs 577 tok/s/user，改动涉及 kernel、通信、路由与调度。
(6/8) On NVIDIA B300, Evo started from vLLM's initial DeepSeek-V4.1-Flash support, with no access to subsequent upstream Git history or web searches. In the final 1-hour AgentX evaluation, Evo surpassed the latest expert-tuned vLLM recipe: • Output throughput: 189.37 vs. 181.85 tok/s (+5%) • P90 interactivity: 664.03 vs. 576.65 tok/s/user (+15%) The improvements span kernels, communication, routing, memory management, and scheduling—not just configuration tuning.
Our AI system, Evo, surpassed the expert-tuned vLLM serving recipe for DeepSeek-V4.1-Flash on NVIDIA B300 in just 3.5 days—starting from vLLM's Day-0 model support, with no human steering and no access to web & history. Coding agents are getting better at solving individual tasks and PRs. With Evo, we're going after something much bigger: Continuously self-improving software systems that automate software engineering. Our thesis: much of software engineering can be reduced to constrained optimization problems. Give Evo a codebase, an objective, and constraints. It decides what to optimize next, scales coding agents to execute, runs experiments in parallel, and keeps only verified improvements. The key is closing the loop between creating tasks and solving them. Every result informs what Evo should attempt next, enabling continuous optimization without human intervention. We started with LLM serving. On DeepSeek-V4.1-Flash with agentic coding workloads (AgentX), Evo outperformed the expert-tuned vLLM recipe on NVIDIA B300: • +15% p90 interactivity • +5% throughput • 5-8× less time to surpass expert performance ML systems engineering is just the beginning.