Whitepaper: Beyond the limits of local LLMs
project veneta whitepaper · October 2026 · v1.3.1
Beyond the limits of local LLMs.
Why the language models an enterprise runs inside its own network answer below expectations, and what changes when a layer of memory and judgment is placed around the model instead of replacing it. Measured on eight open models, 12B to 70B, inside one network.
No form, no account, nothing stored. The PDF opens directly. Two editions: English, 29 pages; Korean, 28 pages. The fairness paragraph and the list of what we do not claim are in the PDF as written.
Where this overlaps with what exists: checks written in code, a critic model and retrieval memory are known patterns, and the paper says so. Where it does not: a lesson is proposed and recalled only after a person approves it, every decision is an undoable ledger line, and the layer is measured on eight open models against a no-memory control.
v1.3.1
What is inside
Five limits
What a local LLM cannot do alone: know the present state, remember, refrain from inventing, check itself, learn under governance.
The layer
The six steps that answer an operational question, the kinds of governed memory, and how checks written in code bind the critic model.
What we measured
Eight open models on one desktop machine: memory on and off, judgment before and after, learned rules applied, and a head-to-head with an open-source memory layer.
Security and governance
Seven probes on the memory as an attack surface: injection at write, secrets, per-user isolation, deletion proof, tamper evidence, rejected lessons, telemetry defaults.
What we do not claim
Synthetic data, one edition, a small experiment, a real latency cost, a person who must approve. The limits are written down before peer review.
Get the whitepaper
Next
Written by VENETA Inc., which maintains project veneta, as the reference measurement for the veneta layer. Authors Hyunjoo Lee and Minjun Lee.
Questions about the measurements: [email protected]