HARDWARE
Detecting hardware…
—
—
—
—
How recommendations are chosen

Recommendations consider machine compatibility, installed Backends, model architecture and format, estimated memory, and available memory now. They prefer official current-family checkpoints and comfortable fits; community and experimental text variants remain in All models.

Recommendation slots use catalog format and tier heuristics, not measured benchmarks. Best quality means the largest fitting tier; fastest means the smallest tier.

Multimodal recommendations show image/audio capabilities and missing add-ons. Install required add-ons in Backends before loading with media. Saved load settings are not changed by the recommendation Backend.

Loading suggestions…
— loaded— downloaded— on disk
INFERENCE
Idle

Loaded models

No model loaded

On disk

Loading models…
Request settings
REQUEST / CONTEXT & QUESTIONS

Questions

RESPONSE / DECISION

Decision

Load a model to enable the playground.

One simple item ({"state", "question", "options", "gold"?}) or one /v1/systemone request per line; one answer per line in the same order (format: docs/SERVING.md#inference-on-a-jsonl-file). Batch requests go straight to the model: they are not counted in the statistics or the history.

Logs

basal-viz loads the bf16 PyTorch weights of basal-1.0, basal-1.5, basal-2.0 or basal-omni into basal-ui’s process (counted in the memory budget) and shows the prompt of every option order, its tokens, the decision after every layer and the calibration. Inspect this decision in the Playground sends the request here and lists what differs from the served decision.

Only this app can reach the model
tokens; read by the vLLM and SGLang modes

API keys

Connect

No output yet.

Load downloaded model

Load a model here, then choose Use when it is ready. Load options and media support are shared with Model manager.

Browse models to download →

Keyboard shortcuts

Ctrl/⌘ + Enter Send Playground decision

? Open this help

Save preset

Test inference

A private model runs one short text decision. No downloads or decision-history records are created.

Optional: unload idle models to free memory

Selected models remain stopped afterward. Use Restore in the test result to reload their previous configuration.