Sayed Amir Karim of Pelloran Inc. submitted a preprint on 10 November 2026 describing tritify, an open-source runtime that, according to its abstract, executes model forward passes entirely in exact integer arithmetic. Floating point remains in offline quantization and the reference used for verification. The reported scope includes decoder-only language models, Mixture-of-Experts (MoE) decoders, and the Segment Anything 2.1 image/video stack.[1]

The abstract separates three claims: bit-identical outputs across runs and hardware, agreement with the floating-point reference’s decisions on tested models and inputs, and unchanged output bits when computation moves between machines. It reports a reproducing SHA-256 output digest and measured numeric exceptions to decision fidelity. These are distinct checks: matching output bytes and matching reference decisions establish different properties.[1]

For distributed execution, the author reports that GLM-4.5-Air, with 106,852,251,264 parameters and 427 GB of integer codes, generated an answer across three machines bit-identically to a single-host reference. None of those machines could hold all its integer codes in RAM. The proposed verification approach uses an open scalar oracle to re-derive exact bytes; the abstract explicitly leaves a guarantee against a dishonest server to future work. It claims neither a per-device speedup nor confidentiality.[1]

[1]: https://arxiv.org/abs/2611.13730

Sources

  1. Deterministic, Reproducible Inference in Exact Integer Arithmetic, and Verifiable Distributed Inference as Its ConsequencearXiv · 10 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

b5411b5152d61e856f6e3e31ac2e020c9a06f61bbd9f510c891b2c714a7d1ea8