About
llmsystems.dev is an operational reference for people who run LLM systems in practice — self-hosted models, local and server runtimes, hardware budgets, and explainable configuration.
It answers questions like: which model variants exist, what artifact sizes imply for VRAM (with honest labels), which runtime is in scope, and what the evidence behind a number actually is.
It is a structured field reference — entities, relationships, evidence-labeled facts, and configuration data — not a news feed or tutorial blog.
What you will find here
- curated model pages and thin artifact records
- a llama.cpp runtime entry
- hardware memory classes and a few representative GPUs
- weights-only VRAM bands (
SOURCE+POLICY), with KV and runtime overhead marked not known yet - methodology pages that define evidence classes
This site does not currently publish full fit results, max-context solvers, or generated startup commands. Where those are not available yet, pages say so explicitly.
Evidence categories
Numbers on this site are labeled:
- SOURCE — from an authoritative artifact or document
- POLICY — an explicit site rule (for example band thresholds)
- OBSERVED — measured, with methodology version and confidence
- CALCULATED / ESTIMATED — when formulas or informed approximations apply
If a value is unknown, the page says so.
See Methodology for the full evidence model.
Who this is for
Primary audience: developers and operators self-hosting or deploying open-weight models with tools such as llama.cpp (and other runtimes as they are added), who need a structured reference rather than another news feed.
Contact
Questions, corrections, or suggestions about the data and methodology: