AXI for Model Hosts: Ten Engineering Tests on Exact State

AXI is a host-side layer between a model and the state it edits. The model is unchanged, and every claim in this paper is about the state the host holds. The paper reports ten engineering questions answered on 11 open models (Tests 35 to 45): the host's state was exact on 80,000 of 80,000 steps on a 7B model (four jobs of 20,000 steps) and in the main cell of 10 of the 11 models; a kept chat computed 1.24 to 4.11 times the prompt tokens per step and ended at its window after 108 to 337 steps; one NVIDIA L4 served 256 sessions with every step exact; none of 1,401,041 injection proposals was committed in a 10,000,000-proposal stress run against the Python reference model; and the focused view shown to the model stayed at a median of 207 to 323 characters while the whole state grew to 378,399 characters. Scope: one NVIDIA L4 on Colab, vLLM 0.31.0, greedy decoding, synthetic service graphs; Tests 39 and 40 ran on a CPU against the Python reference model, not the Rust crates; the 32B ran with part of its weights offloaded to the CPU. The framework and architecture are by Rajnish Choubey; the tests were built and run by Shalendra Gaur. This paper continues the earlier white paper "AXI: The Cure for the Cost Crisis in AI" (https://doi.org/10.5281/zenodo.23958737). Code, test pages and result files: https://github.com/gaurshalendra2412/axi-runtime

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24319560
Primary Topic
Natural Language Processing Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

AXI for Model Hosts: Ten Engineering Tests on Exact State

Architect Rajnish choubey, Shalendra Gaur
3 links
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
preprint

AXI for Model Hosts: Ten Engineering Tests on Exact State

Architect Rajnish choubey, Shalendra Gaur
preprint en

Abstract

AXI is a host-side layer between a model and the state it edits. The model is unchanged, and every claim in this paper is about the state the host holds. The paper reports ten engineering questions answered on 11 open models (Tests 35 to 45): the host's state was exact on 80,000 of 80,000 steps on a 7B model (four jobs of 20,000 steps) and in the main cell of 10 of the 11 models; a kept chat computed 1.24 to 4.11 times the prompt tokens per step and ended at its window after 108 to 337 steps; one NVIDIA L4 served 256 sessions with every step exact; none of 1,401,041 injection proposals was committed in a 10,000,000-proposal stress run against the Python reference model; and the focused view shown to the model stayed at a median of 207 to 323 characters while the whole state grew to 378,399 characters. Scope: one NVIDIA L4 on Colab, vLLM 0.31.0, greedy decoding, synthetic service graphs; Tests 39 and 40 ran on a CPU against the Python reference model, not the Rust crates; the 32B ran with part of its weights offloaded to the CPU. The framework and architecture are by Rajnish Choubey; the tests were built and run by Shalendra Gaur. This paper continues the earlier white paper "AXI: The Cure for the Cost Crisis in AI" (https://doi.org/10.5281/zenodo.23958737). Code, test pages and result files: https://github.com/gaurshalendra2412/axi-runtime

Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon