457 points on Hacker News, 268 comments, top technical post of the day. Google published HEIR — Homomorphic Encryption Intermediate Representation — on August 14, 2026, and the pitch is one word: practical.
What it actually is
Not a model, not an API service. HEIR is an MLIR-based open-source compiler toolchain. You feed it a pre-trained model, it emits a version that runs inference directly on encrypted inputs. No decrypt-then-compute step on the server side. Google demoed it on a deep learning recommender, credit card fraud scoring, Kitsune network intrusion detection, and hotword detection for voice agents.
Fully homomorphic encryption has been a paper toy for a decade — the tax was 1000x slowdown. Google’s own numbers now: a 3-layer TFLite net compiled to FHE, private inference in 16 seconds single-threaded CPU. Still slow. But fraud scoring lives in a 1–10 second window, and hardware partners Belfort, Niobium, Cornami and Optalysys are closing that gap.
Getting in
Apache-2.0 on GitHub, 812 stars. Backends: OpenFHE, Lattigo, tfhe-rs, Jaxite. Schemes: BGV, BFV, CKKS. You wire it into your own model conversion pipeline — the point is interoperability across schemes, compilers and accelerators, so nobody hand-writes FHE circuits again. Four peer-reviewed papers already built on it, with Georgia Tech, CMU, Purdue and Tsinghua.
You Might Also Like
- Google Deepmind Aletheia Just Solved Math Problems Nobody Could Heres why That Matters
- Google Workspace Studio Just Made ai Agents a Thing Everyone can Build
- Google A2ui Agent to User Interface Finally a Standard way for ai Agents to Show you Things
- Google Lyria 3 Just Turned Gemini Into a Music Studio and im Weirdly Into it
- Google Pomelli Photoshoot Just Dropped and Product Photographers Should be Paying Attention
Top AI Product 
Leave a comment