Meta's Pyrefly Reaches 1.0, Bringing a Rust-Built Python Type Checker to Stable With 90% Spec Conformance
Pyrefly 1.0.0 ships on May 12, 2026 as Meta's production-grade Python type checker, checking 1.8 million lines per second and powering Instagram's codebase.
Overview
Meta released Pyrefly 1.0.0 on May 12, 2026, marking the graduation of its Rust-built Python type checker from beta to a stable, production-ready release. The tool processes 1.8 million lines of code per second, according to Engineering at Meta, and is already the default type checker for developers working on Instagram’s Python codebase.
What We Know
Pyrefly is Meta’s successor to Pyre, an OCaml-based type checker originally developed in 2017 to serve Instagram’s typed Python codebase. According to Engineering at Meta, the team built a new engine from the ground up in Rust rather than extending the older tool, using a custom incremental computation model and an IDE-first design that shares a consistent view between editor integrations and command-line use.
The 1.0.0 release ships several concrete improvements over the November 2025 beta, according to the official release notes:
- Conformance to the Python typing specification improved from 70% at beta to over 90% at stable release
- 2–125x faster updated diagnostics after file saves
- 20–36% faster full type checking on large projects
- 2–3x faster initial indexing
- 40–60% less memory usage during indexing and incremental checking
- Over 60 minor releases shipped since beta, fixing hundreds of bugs
The release follows over six months of beta testing since November 2025, and follows the alpha release in May 2025 when InfoQ reported that Pyrefly could check PyTorch’s codebase in 2.4 seconds compared to 35.2 seconds for Pyright and 48.1 seconds for Mypy at the time.
Architecture
Pyrefly’s speed advantage traces to its incremental computation model, described by Engineering at Meta as built around modules as the fundamental unit of type checking. The tool uses the module’s public interface to determine whether downstream modules need rechecking — if the interface does not change after an internal edit, no downstream work is triggered. This design makes it possible to deliver responsive checking in editors while still handling monorepo-scale codebases.
The tool is written in Rust and uses multi-threaded parallel checking across available CPU cores. The Engineering at Meta post notes that Rust was chosen for its type safety, threading support, and performance characteristics.
New Features in 1.0.0
The stable release introduces configuration presets labeled “off,” “basic,” “legacy,” “default,” and “strict,” reducing setup friction for new users. Projects without a configuration file default to the basic preset, which surfaces only high-confidence errors likely to indicate real bugs, according to the release notes.
Automatic configuration synthesis from existing mypy or pyright configuration files is included, enabling most projects to migrate with reduced manual work, according to the release notes.
Framework support was expanded significantly. The 1.0.0 release adds Django model relationship awareness and factory_boy factory support, Pydantic lax-mode validation, Pytest code lens and fixture annotations, and Jupyter notebook parity with standard .py files. Experimental support for tracking tensor dimensions through PyTorch models is also included, according to the release notes.
For IDE users, the release ships Safe Delete refactoring, bulk fixAll operations, and improved workspace stability. The release notes report that updated diagnostics after file saves are 2–125x faster compared to prior versions.
The toolchain includes pyrefly infer for bootstrapping type annotations in unannotated code, pyrefly init for migrating configurations from mypy or pyright, and pyrefly coverage report for annotation-completeness metrics.
Adoption and Ecosystem Integration
Pyrefly is already deployed in production at several large Python codebases. The release notes confirm that Pyrefly was already the default type checker for developers working on Instagram at Meta and has been adopted by PyTorch and JAX. The tool is MIT licensed and available via pip install pyrefly.
JetBrains integrated Pyrefly as an optional type provider in PyCharm 2026.1.2, described in the PyCharm blog as “Meta’s next-generation Python type checker, engineered from the ground up in Rust to replace its predecessor, Pyre (written in OCaml).” The integration resolves files in 0.5–1 seconds, and is accessible through a Type widget in the IDE status bar. JetBrains notes that support for Docker, Docker Compose, WSL, SSH, and multi-module projects is planned for future releases.
What We Don’t Know
Pyrefly’s conformance score, while improved to over 90%, remains lower than Pyright’s, and the 1.0.0 release notes do not publish the precise current gap. The PyTorch tensor dimension tracking feature is marked experimental with no timeline for stabilization. The tool has not yet been benchmarked against current versions of Pyright and Mypy in a published comparison from an independent source.
Pyrefly enters a competitive Python type-checking landscape alongside Astral’s ty. As InfoWorld notes, Pyrefly offers a “broader existing feature set, better documentation, and tooling to allow elegant migration,” while ty is noted for its “contextually detailed error messages.” Whether the two Rust-based tools will consolidate or further fragment the type-checking ecosystem remains an open question.