Trio-Spark v1.3 adds /v1/systemone and a vendor-run panel with calibration, and discloses two regressions
MachineFi Labs released Trio-Spark v1.3 on 15 October 2026. The listing on this site was out of date, and two vendor-submitted corrections (the submitter, an IoTeX cofounder, disclosed his MachineFi affiliation again) are now reflected on the Trio-Spark page.
What changed, checked at the source. GET /v1/models lists trio-spark-v1.0 to v1.3, with v1.3 dated 2026-10-15, text, image and sampled-video input, a 1,024-token context and the same $0.042 per million input tokens. The vendor’s visual guide documents POST /v1/systemone for trio-spark-v1.3, and the endpoint answers 401 without a key. /v1/models also lists noul and score, but the public docs only show Choice examples, and we haven’t run requests ourselves.
Numbers (UnverifiedClaim). On a ten-task panel of 2,977 decisions the vendor reports 90.9% for v1.3, against 83.1% for v1.0, 83.0% for v1.2 and 90.0% for Jev 1.13 on the same items (Clef 90.8%, Clef-Flash 91.8%). Expected calibration error is 0.019, against 0.058 for v1.2 and 0.234 for v1.0. The vendor discloses that v1.3’s training data includes the public training splits of those tasks, so this is not a held-out test, and that rejected requests count as wrong. It also says v1.3 adds about 85 ms per request and scores below v1.0 on an internal GUI-action suite.
Still missing. No Decision Index or JevBench score yet: the vendor says it submitted v1.3 to JevBench through benchmarkheaven.com on 15 October (regular queue). S1MB closed its request on 13 October after declining a hosted-API evaluation.
Scores. Trio-Spark moves from 4.2 / 3.7 / 5 to 4.9 maturity / 5.9 capability / 5 adoption: /v1/systemone compatibility and a fuller guide lift maturity, and the three listed decision types plus a published ECE lift capability. It stays well below Jev, and the drone and JevBench latency figures predate v1.3.