MachineFi's Trio-Spark joins the catalog as a beta Choice API, with its lineage undisclosed
MachineFi Labs, the IoTeX-affiliated lab behind the Trio products, launched Trio-Spark v1.0 on 30 September 2026 (X post at 11:12 BRT). It is a hosted decision API: you send a task, the current state and 2–8 allowed actions, and it returns one choice_id, a probability for every option, a confidence value and a decide / review / abstain disposition, with no generated text. Pricing is $0.042 per million input tokens with no output charge, and new accounts get 100 free decisions.
How it got here. Raullen Chai, IoTeX cofounder, submitted it through hub issue #2 and disclosed his affiliation. After our first review, MachineFi published an anonymous API reference covering schema, limits and errors, and answered our lineage questions in the issue.
What we know, and what we don’t. The endpoint is POST https://platform.machinefi.com/api/spark/v1/decisions with model id trio-spark-preview. It is Choice only and not /v1/systemone-compatible, so the TypeSafe SDK doesn’t work with it. Weights, base model, size and training method are not disclosed. MachineFi states that no request calls TypeSafe Jev or another third-party decision model at inference time, which we can’t verify. The price is identical to Jev’s, and to the price at which QuickSilver Pro, a service by the same founder, resells Jev. MachineFi also withdrew a “sub-100ms” launch claim: its drone demo shows 933 ms median end-to-end wall time.
Numbers (UnverifiedClaim). In a vendor-run JevBench preflight on the public easy + original items, Trio-Spark got 116/120 (easy 47/48, original 69/72) at ~0.51 s median wall time. Long hard items hit a gateway 503. MachineFi calls this a preflight, not a leaderboard score. Official evaluation is requested at JevBench #166 (adapter PR #165) and S1MB #1.
How we list it. Trio-Spark enters the catalog as beta at 4.2 maturity / 3.7 capability / 5 adoption, with a caveat box at the top of the page. That puts it well below Jev (6.0 / 8.2 / 58) and d1 (4.9 / 6.5 / 23), mainly because it supports one decision type and has no independent numbers yet. We’ll revisit the scores when JevBench or S1MB publish results.
Correction, 15 Oct 2026: this post originally quoted a compute-only latency figure and the serving hardware behind the launch numbers. The vendor says it no longer publishes serving hardware and that it has changed since launch, so both were removed; the 933 ms wall-time figure stands. The scores above were the 30 Sep values. Trio-Spark has since moved to v1.3 and 4.9 / 5.9 / 5 (see Trio-Spark v1.3).