Google showcased an open-source compiler called HEIR on August 14, 2026, as part of its Private Computing Toolkit, presenting it as a way to run artificial-intelligence inference without exposing the underlying input data to the server performing the calculation. HEIR stands for Homomorphic Encryption Intermediate Representation.
Homomorphic encryption allows operations to be applied directly to ciphertext. A service can receive encrypted input, compute an encrypted result and return it without first viewing the original information. Google offered private content recommendations as one example: a cloud model could process protected user features without learning those features in readable form.
That approach addresses a limitation of ordinary end-to-end encryption. Encryption can shield stored or transmitted information, but a conventional server usually needs plaintext to perform tasks such as classification. Running the model locally avoids sharing data but may exceed the device’s capacity, and distributing a proprietary model can expose the provider’s intellectual property. Homomorphic encryption replaces part of that capability-versus-privacy trade-off with computational cost.
HEIR is intended to reduce the specialist work required to make encrypted computation practical. Google says the compiler can transform pre-trained models designed for ordinary data into versions that accept encrypted inputs. The longer-term goal is a one-click workflow through which developers without cryptography expertise can add encrypted inference to production applications. The current announcement presents that as a vision rather than a completed guarantee for every model.
Google first outlined its plans for HEIR in 2023. Since then, the project has been used as a research platform and has supported four peer-reviewed publications, with additional work in preparation, according to the company. Collaborators named in the announcement include Georgia Tech, Carnegie Mellon, UC Santa Barbara, Illinois Institute of Technology, Purdue, the University of Edinburgh and Tsinghua University.
The project is also working with homomorphic-encryption hardware companies Belfort, Niobium, Cornami and Optalysys. Google said accelerator demonstrations were part of its work and that it planned to show latency improvements in the future. That future wording is important: the supplied evidence does not provide comparative accelerator results or establish production economics.
Four private-inference applications were compiled with HEIR and released with source code, with latency reported for a single-threaded CPU. The excerpt does not include the individual measurements, so no performance comparison can be drawn from it. Google acknowledges that homomorphic encryption still imposes non-trivial overhead even as costs decline.
HEIR therefore addresses developer accessibility more directly than it resolves every performance constraint. Its open toolchain offers common compilation, testing and benchmarking infrastructure that researchers can reuse. Whether encrypted inference becomes routine will depend on model compatibility, latency, hardware support and operational cost, but the project provides a concrete route for evaluating AI services that keep inputs encrypted throughout server-side computation.
