Meta built its own in-house Copilot

Meta is tired of watching Microsoft and GitHub have all the fun with AI-assisted coding, so it built its own version behind closed doors.

At an internal AI infrastructure event, the tech giant unveiled CodeCompose, an in-house generative AI tool that acts as Meta's private alternative to GitHub Copilot. For now, it is not available to the public and is strictly reserved for internal engineering teams.

Meta software engineer Michael Bolin shared that developers across the company are already using CodeCompose inside IDEs like VS Code, getting real-time code completions for Python and other languages as they type.

Half the size of Copilot, but tuned for Hack

The largest version of CodeCompose tops out at 6.7 billion parameters. That is just over half the size of the underlying model powering GitHub Copilot, but Meta is leaning into specialization rather than sheer parameter count.

CodeCompose is built on top of InCoder, an open research model previously released by Meta AI, and then fine-tuned heavily on Meta's internal codebases. This includes proprietary libraries and frameworks written in Hack, Meta's home-grown programming language that public models know very little about.

Meta also filtered its training data to weed out bad coding habits and deprecated APIs before feeding it into the model. In everyday use, it suggests code annotations, fills in import statements, and generates multi-line blocks based on surrounding code and comments.

The tool is already seeing heavy traction. Meta claims thousands of its employees accept suggestions from CodeCompose each week, with an acceptance rate topping 20%.

Meta is also trying hard to avoid the legal crosshairs currently aimed at GitHub and OpenAI, who are facing a class-action lawsuit over Copilot reproducing copyrighted open-source code without credit. According to a Meta spokesperson, InCoder was strictly trained on Stack Overflow and public repositories with permissive, non-copyleft open-source licenses, with the only additional training coming from Meta's own proprietary code.

Why it matters

General-purpose coding models are great for boilerplate, but they fall apart the moment a company relies on bespoke internal infrastructure. Meta's approach proves that a leaner 6.7-billion-parameter model, tailored to a private codebase and weird internal languages like Hack, can deliver real utility without leaking company IP to third-party APIs.

We think keeping this in-house is a clever defensive move. With a 20% acceptance rate, CodeCompose is actually delivering utility rather than sitting as internal shelfware. It keeps Meta developers fast while shielding proprietary logic from external cloud providers.

The catch is long-term code health. As recent Stanford research highlighted, developers who lean on AI assistants frequently introduce more security vulnerabilities into their builds. If CodeCompose ends up learning subtle legacy bugs or bad architectural patterns from Meta's existing codebase, it could end up automating technical debt at unprecedented speed.