Voice-LLM Bench / AutoResearch

Nightly tool-calling latency trials against the Ollama Cloud proxy · winner is promoted to the voice-agent default
last run —
window —
experiments —

Latency by model

Median of 3 tool-calling turns, successful batches only — latest night vs. full window

Latency trend

Per-night median; gaps are failed nights. Click legend to isolate a model.

Reliability

Share of nights with 3/3 valid tool calls, plus raw error count

Latency heatmap

Weekly median per model — greener is faster, grey = not sampled

fastmidslowno data

Who held the crown

Fastest valid model per night; a new segment means the lead changed hands

Model scoreboard

Click a header to sort. Median is the median of nightly medians — the fair cross-night number.