Home›Companies›Relace

Relace

LLMs for Code Generation

Relace makes it easy for startups to productionize projects that involve coding agents, with specialized LLMs for retrieval and code merging. Our SOTA apply model merges code at 2500+ tok/s, and our code-specific reranker & embedding models can process million-line codebases in <1s.
Active Founders
Preston Zhou
Preston Zhou

Preston Zhou, Founder

Physics -> ML
Eitan Borgnia
Eitan Borgnia

Eitan Borgnia, Founder

Former PhD student in machine learning at UChicago, with a math degree from Caltech.
Jobs at Relace
San Francisco, CA, US
$100K - $200K
Any (new grads ok)
Relace
Founded:2022
Batch:Winter 2023
Team Size:3
Status:
Active
Primary Partner:Brad Flora
Company Launches
Relace - Cheaper, faster, and more reliable AI code generation.
See original launch post

Hey everyone 👋 -- we're Preston and Eitan, the cofounders of Relace.

TL;DR:

Relace models are designed to slot naturally into most AI codegen products, making them faster, cheaper, and more reliable. We already have SOTA:

  • Embedding + Code Reranker models to retreive the relevant context from million-line codebases in ~1-2s.
  • Instant Apply model that merges code snippets at >4300 tok/s.

We're in production with Lovable, Magic Patterns, Codebuff, Create, Tempo Labs, and 20+ other AI codegen startups. Watch our Instant Apply model race against full file rewriting with Claude 3.7 Sonnet:

https://www.youtube.com/watch?v=J0-oYyozUZw

Problem:

Agentic coding systems are easy to prototype, but hard to make robust. As users succeed at creating more complex designs/applications, you hit bottlenecks:

  • Larger codebases require efficient context management -- you can't pass everything to the AI agent.
  • Slow and expensive full file rewrites need to be replaced with abbreviated diff formats.

Frontier models (like Claude 3.7, o3, etc.) are powerful, but overkill for these auxiliary functions like retrieval and merging. Costs start to add up, especially when every agent action hits an API with thousands of tokens in and out. Plus, non-technical users are easily frustrated by high latency, especially if they are trying to refine small aspects of the code they created.

Solution:

Relace models are trained to achieve order-of-magnitude improvements in both latency and cost without compromising on accuracy.

Instant Apply

Released in February, this model merges semantic code snippets at 4,300 tokens/sec with an average end-to-end latency of ~900ms across our users.

Inspired by Cursor's Fast Apply, the semantic diff format is chosen to be natural for all models to output. A simple prompt change combined with Instant Apply can reduce Claude 3.5/3.7 token usage by ~40%.

We train on hundreds of thousands of code merges across dozens of different coding languages to achieve SOTA accuracy:

Embeddings and Reranker

Our embedding model + reranker can determine the relevance score for a user request against million-line codebases in ~1-2s. By training on hundreds of thousands of query/code pairs, we can effectively filter out irrelevant files and cut input token usage by over 50%.

Not only does this save on cost, but cleaning up the context window significantly improves generation quality of the AI agent.

Try It Out:

Both of these models are battle tested and running millions of times a week in production. You can read more and try it out for yourself with the links below.

App: https://app.relace.ai Docs: https://docs.relace.ai

We have a free tier for small projects, and we offer discounted rates for open source partners like our friends at Continue.

Don't hesitate to reach out if you're optimizing your coding agent — we would love to hear your thoughts, feedback, and what you're building!

Other Company Launches

Squack - Natural language RPA tools for accountants

We let you streamline your data workflow by building tools with everyday language.
Read Launch ›