tritify
The same bytes on every machine: real models, run in exact integer arithmetic.
The code on GitHub ↗Read the preprint (PDF)The published preprint (DOI) ↗
What it does
tritify runs the whole forward pass of real models in exact integer arithmetic: language models, mixture-of-experts models, and an image and video model. Its output is bit-identical run to run and machine to machine, carried by a SHA-256 hash that you can reproduce and check yourself.
Why it matters
Optimized floating-point inference is usually not reproducible: the same model on different hardware, compilers or thread counts can produce different bits.
What we claim, and you can check
- Reproducible across hardware: the same output bits on every run, across processor architectures, and across CPU and GPU.
- Faithful to the original: on the models and inputs tested it makes the floating-point model's decisions, and the exceptions are measured, not assumed away.
- Independent of where it runs: a model's computation can be split across processors and across machines without changing a single output bit.
Shown at scale
We showed the third claim at scale: GLM-4.5-Air, a model with 106 billion parameters and 427 GB of integer codes, generated its answer across three machines, none of which can hold all its codes in memory, bit for bit the same as a single machine.
So you do not have to take a fast fleet on trust: an open reference recomputes the exact bytes and checks them.
What we do not claim
It is not faster on a single device: spreading a model across machines adds latency, and the gain is capacity. It does not keep data confidential: verifiable is not private. And turning agreement between honest runs into a guarantee against a dishonest server is still future work.