Relace

Relace is an AI infrastructure company building specialized models for coding agents — developing apply models (that precisely integrate AI-generated code changes into existing codebases), embedding models optimized for code search and semantic retrieval, and reranking models that filter AI coding agent outputs for quality. Founded and headquartered in San Francisco, Relace raised $23 million in a Series A led by Andreessen Horowitz in October 2024, serving AI coding platform customers including Lovable and Magic Patterns with 1-2 second codebase context retrieval and 10,000+ tokens per second merge speed.\n\nRelace's models address the specific technical challenges of autonomous coding agents that general-purpose LLMs handle poorly — applying code diffs precisely without introducing formatting errors, searching large codebases semantically to find relevant context without overwhelming the model's context window, and filtering generated code for quality and correctness before applying changes. These specialized inference capabilities enable coding agents to work accurately on real production codebases where precision matters, rather than just generating plausible-looking code that fails in context.\n\nIn 2025, Relace operates in the AI coding infrastructure market alongside the models and tools that power the rapidly growing autonomous coding agent category — including Cursor, GitHub Copilot, and AI-native development platforms like Lovable. The apply model is a specific technical capability that multiple coding platforms need: when an LLM suggests a code change, reliably applying that change to the correct location in the file without corrupting surrounding code is harder than it appears. Relace's specialized inference layer enables coding agent companies to achieve higher accuracy without building custom models. The Andreessen Horowitz Series A validates the infrastructure opportunity in the AI coding stack. The 2025 strategy focuses on growing the customer base among AI coding platforms, improving merge accuracy benchmarks, and expanding the model suite to cover more coding agent workflow requirements.