The 100 ms claim held for medians and for the 90th percentile once tool vectors are cached, but not for the first request with a new tool list, so state the median.
Relay 1 (21.8M parameters, 11.47 MB with 4-bit weights) as .gust graphs for the Gust runtime, with the tokenizer, the optional retrieval prefilter and checksums. Weights are CC BY 4.0.
