Settings and defaults

Current defaults

The source defaults are Lemonade, ordinary layer placement, zero explicit CPU threads, Night, 8,192 context, 12 GPU layers, CUDA, temperature 0.7, and 512 maximum output tokens. The experimental Strata cache default is 2,048 MiB; selecting the engine is explicit. Your saved preferences take precedence. These are starting values, not artifact-specific performance recommendations.

Open System to change sampling and appearance. Save preferences saves temperature, output limit, theme and the current Load Lab engine, context, placement, GPU-layer policy, cache and thread options as defaults. Changing a saved load value does not retroactively resize an already loaded model; reload deliberately when a new allocation configuration is needed.

Setting	Current accepted range
Temperature	Finite number, 0–2
Maximum output	Lemonade 1–32,768; Strata 1–4,096 tokens
Context	Lemonade 128–262,144; Strata 256–8,192; constrained further by model metadata
Backend	Lemonade CUDA or CPU; initial Strata CUDA only
CPU threads	Lemonade0 or1 through logical-core limit; Strata0 or2 through that limit, counted as total participants
Expert cache	Strata512–4,096 MiB, step256; irrelevant to Lemonade
Appearance	Night or Day

An existing value is never silently shortened when an engine is selected. Out-of-range values remain visible and invalid; the active host adapter also revalidates requests independently from the UI draft.

The output limit is an upper bound. A stop condition, error, cancellation or model completion may end generation sooner. A low temperature does not by itself guarantee determinism across kernels, builds and hardware.

Appearance and animation

The native app has Night and Day modes. Its startup nebula is full-window; About has a bounded floating presentation. Both honor reduced-motion and visibility behavior, and release GPU resources when unmounted. Chat uses static identity art, so the decorative scene does not continue running there.

The documentation site’s Day/Night choice is separate from native settings and is stored in browser localStorage as regesarc-docs-theme. Starlight also retains sidebar navigation state in the current browser session. Changing a docs theme does not alter a desktop load plan or inference state. No hosted analytics or search service is configured.

Controls still planned

Named presets, per-model defaults, a system-prompt editor, Jinja template editing, top-p/top-k/min-p, repeat penalties, KV-cache precision, flash attention, batch-size and advanced thread-affinity controls, speculative decoding and engine-argument editors are not current native controls. Some are upstream capabilities, but that is not an implemented RegesARC contract.

Do not edit internal runtime configuration to simulate a supported UI setting during qualification. Protected context, model path, port, credentials and placement options must retain one owner. Future advanced controls need validation, a visible effective configuration and saved-version migration. See the roadmap for those gates.