Trio-Spark

Last updated: October 15, 2026

Beta510–933ms$0.042/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

Maturity4.9/10
Self-serve hosted API with versioned model ids (v1.0–v1.3) at a public /v1/models, a REST reference and a visual/System One guide. Now serves /v1/systemone, so System One-style requests apply. No third-party integrations, no SLA; the platform docs sit behind a login.
Capability5.9/10
v1.3 lists Choice, Score and Noul (the public docs only show Choice examples) plus image and sampled-video input. Vendor ECE 0.019 and 90.9% on a ten-task panel it trained on, vs Jev 90.0%. Wall time 0.5–0.9 s, +85 ms in v1.3. Everything is vendor-run; no Decision Index or JevBench result.
Adoption5/100
Launched 30 Sep 2026. Main launch post ~6 likes / 399 views; founder and IoTeX quote posts ~240 and ~185 likes. Repo 3 stars, HF card 0 downloads. No HN or press coverage, no integrations or production cases. Not re-measured for v1.3.

Vendor claims

v1.3 (released 2026-10-15) answered 90.9% of a ten-task panel of 2,977 decisions correctly, vs 83.1% for v1.0 and 83.0% for v1.2; on the same items Clef 90.8%, Jev 1.13 90.0%, Clef-Flash 91.8%. Peers were run through their official APIs/SDK on 7–8 Oct 2026; vendor-run, not independently reproduced[Vendor claim — not independently verified]
Expected calibration error (10 bins) on that ten-task panel: 0.019 for v1.3, 0.058 for v1.2, 0.234 for v1.0[Vendor claim — not independently verified]
v1.3 training data includes the public training splits of the panel tasks, no evaluation item was used for training; served 4-bit weights; requests the API rejects (over 1,024 input tokens, or content-filtered) count as wrong; Banking77 (77 options) is read out offline from the same weights[Vendor claim — not independently verified]
Disclosed regressions: about +85 ms per request vs v1.2, and a lower score than v1.0 on an internal GUI-action suite[Vendor claim — not independently verified]
GET /v1/models lists trio-spark-v1.0 to v1.3; v1.3 has release_date 2026-10-15, supported_features boolean/choice/noul/score, input_modalities text/image/video_sampled_frames, context_length 1024, $0.042 per million input tokens and no output charge[Vendor claim — not independently verified]
POST /v1/systemone serves trio-spark-v1.3 for text, image (one JPEG/PNG) and sampled-video (2–4 frames, at most 30 s) input; the native /api/spark/v1/decisions endpoint is unchanged[Vendor claim — not independently verified]
Launch accuracy on the production API (v1.0): SST-2 92.89%, AG News 88.46%, TweetEval offensive 80.58% (PubMedQA 85.33% on the HF card)[Vendor claim — not independently verified]
The production API runs Trio-Spark directly and does not call TypeSafe Jev or another third-party decision model during inference[Vendor claim — not independently verified]
JevBench public preflight on easy + original items (v1.0): 116/120 correct (easy 47/48, original 69/72), 120/120 valid distributions, median wall latency ~0.51 s; long hard items hit the gateway's retryable 503. A preflight, not a leaderboard score[Vendor claim — not independently verified]
Drone demo (v1.0): 25 decisions, 0 collisions, 933 ms median / 1,049 ms p90 end-to-end wall latency[Vendor claim — not independently verified]

Read this first

Lineage is undisclosed. MachineFi says it trained and operates Trio-Spark but is not publishing the base model, parameter count, weights or training method. The vendor states that no request calls TypeSafe Jev or another third-party decision model at inference time (API reference, hub reply). That is a vendor claim we cannot verify.

Same price as Jev resale. $0.042 per million input tokens with free output is exactly Jev 1.13’s list price, and the price at which QuickSilver Pro, a service by the same founder, resells Jev. MachineFi calls it “intentional category-reference launch pricing”.

Vendor-submitted. This listing came in through hub issue #2, and the v1.3 update through issues #3 and #4. All three were opened by Raullen Chai (IoTeX cofounder), who disclosed his affiliation with MachineFi.

The “sub-100ms” claim was withdrawn. It appeared in a launch post. MachineFi now says the correct public figure is the drone run’s 933 ms median end-to-end wall time (v1.0). v1.3 adds roughly 85 ms per request over v1.2, by the vendor’s own account.

Every number on this page is vendor-reported (UnverifiedClaim), including the v1.3 panel. The vendor says v1.3’s training data includes the public training splits of those tasks. S1MB #1 was closed on 13/10: the maintainer declined on 01/10 to evaluate the hosted API and offered to revisit if weights are released. JevBench: the vendor says v1.3 was submitted on 15/10 through benchmarkheaven.com (reference BBAEFC3A, regular queue), as the maintainer asked in #166. There is no official score yet.

Trio-Spark (current release v1.3, 15 October 2026) is a hosted decision model from MachineFi Labs, the IoTeX-affiliated lab behind the Trio video and world-model products. MachineFi calls it the first “Situated World Model”: you send the current state and the 2–8 moves your application allows, and one pass returns the selected move and a probability for every option, with no generated text. v1.0 launched on 30 September 2026 (X post at 11:12 BRT); v1.1 added image and sampled-video input the same day, v1.2 followed on 8 October and v1.3 on 15 October.

Specs

Attribute	Value
Company	MachineFi Labs
Launch	30 Sep 2026 (11:12 BRT); current release v1.3 on 15 Oct 2026
Status	Public self-serve beta. Model ids `trio-spark-v1.0` to `trio-spark-v1.3` (the native endpoint keeps `trio-spark-preview` for existing clients)
Endpoints	`POST https://platform.machinefi.com/v1/systemone` (primary) and the native `POST https://platform.machinefi.com/api/spark/v1/decisions` (unchanged)
Auth	Bearer key (`tf_…`) from the platform console, plus a required `Idempotency-Key` header
Decision types	The vendor’s `/v1/models` lists `choice`, `noul` and `score` (and `boolean`) for every version. The public docs only give Choice examples (2–8 alternatives per question, up to 16 questions per System One request). Noul and Score have no documented example
Input	Text (`state` plus typed `questions`), one JPEG/PNG image, or a 2–4 frame sampled-video window of at most 30 s (v1.1 and later). At most 1,024 model input tokens including visual input, rejected rather than truncated. The API does not take MP4, RTSP or livestream URLs
Output	A `choice` and `probabilities` per question on System One; the native endpoint also returns `confidence` and `disposition` (`decide` / `review` / `abstain`)
`/v1/systemone` compatible	Yes, since v1.3 (the vendor calls it a supported subset of the System One image request shape). The native endpoint keeps its own shape
Pricing	$0.042 per million billed input tokens, no output-token charge (media tokens are billed as input), 100 free decisions for new accounts, $5 minimum top-up
Weights	Not published. The HF repo holds only a model card
Parameters / base model	Not disclosed (the v1.3 post says the served weights are 4-bit)
Latency	Vendor wall time 0.51 s median (v1.0 JevBench preflight) to 0.93 s median (v1.0 drone demo); v1.3 adds ~85 ms vs v1.2; no p99

API shape

The native text endpoint, unchanged since v1.0:

curl https://platform.machinefi.com/api/spark/v1/decisions \
  -H "Authorization: Bearer $TRIO_SPARK_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{
    "model": "trio-spark-preview",
    "task": "Keep the machine safe while completing the operation.",
    "state": {"temperature_c": 84, "load_pct": 91, "vibration": "rising"},
    "choices": [
      {"id": "continue", "description": "Continue at the current speed"},
      {"id": "slow", "description": "Reduce speed and keep observing"},
      {"id": "stop", "description": "Stop the machine now"}
    ]
  }'

The response names one choice_id and returns the full distribution in request order, with a disposition that lets the caller route review and abstain cases to a person or a fallback.

The System One form, from the vendor’s visual guide, takes model: "trio-spark-v1.3", a state, typed questions with criteria, and optionally images or media, and answers each question with a choice and probabilities keyed by your criteria. Unlike the native endpoint, it can carry several questions about one state in a single call.

Trio-Spark v1.3 (15 October 2026)

MachineFi’s v1.3 post ships new text weights behind the same API, price and limits. Everything below is a vendor claim (UnverifiedClaim).

Ten-task panel, 2,977 decisions: v1.3 answers 90.9% correctly, up from 83.1% for v1.0 and 83.0% for v1.2, and better than v1.0 on every task where both were measured. On the same items, Clef scores 90.8%, Jev 1.13 90.0% and Clef-Flash 91.8%. The vendor’s own summary is mixed: v1.3 has the best score of the seven models on PubMedQA, RTE, TweetEval offensive and SNLI; others lead elsewhere (Clef-Flash on six rows including HellaSwag, WinoGrande and Banking77, Jev on PIQA, Clef on BoolQ, Laya on AG News). The post’s table has 13 rows; the 90.9% is the ten-task panel total, not a 13-task total.
Calibration: expected calibration error (10 bins) on that panel is 0.019, against 0.058 for v1.2 and 0.234 for v1.0.
How it was measured: the production API code with the served 4-bit weights, one request per item, peers through their official APIs or SDK on 7–8 October 2026. Requests the API rejects (over 1,024 tokens, or content-filtered) count as wrong. v1.3’s training data includes the public training splits of these tasks (no evaluation item was used for training), so the panel is not a held-out generalisation test.
Disclosed regressions: about +85 ms per request versus v1.2, and a lower score than v1.0 on an internal GUI-action suite. The vendor suggests measuring on your own screens before raising automation thresholds if your agent drives a GUI.
Verified at the source: GET /v1/models lists trio-spark-v1.0 to v1.3, with v1.3 at release_date: 2026-10-15, text/image/sampled-video input and a 1,024-token context. POST /v1/systemone answers 401 without a key, so the endpoint exists. We did not run requests against it (no key), so Noul and Score behaviour is not tested by us.

Evidence so far (UnverifiedClaim)

v1.3 panel and ECE: see the section above. Vendor-run, partly on tasks the model was trained on.
JevBench preflight (v1.0) (issue #166): MachineFi ran the public easy + original items through its own adapter and reports 116/120 correct (easy 47/48, original 69/72), 120/120 valid distributions and ~0.51 s median wall latency. The first long-policy hard items hit the production gateway’s retryable 503, and the run stopped there. MachineFi itself says this is a preflight, not a leaderboard score. Raw responses are private.
v1.0 launch post: SST-2 92.89%, AG News 88.46%, TweetEval offensive 80.58% on the production API, with PubMedQA 85.33% on the HF card. The launch post on X adds more rows, some of them “coverage” rather than accuracy.
Demos (repo): a drone (25 decisions, 0 collisions), a browser agent (3 calls, 4.0 s) and a robot arm (13 calls, physical success). Each one is a patch over an MIT-licensed Jev demo with the TypeSafe call swapped for Spark. THIRD_PARTY.md credits the three upstreams: RomanSlack/jev-drone, browser-use/jev-ultrafast and FBddcz/embodied-jev.
Not on the official Decision Index or the JevBench leaderboard as of 15 Oct 2026. S1MB declined a hosted-API evaluation (01/10) and closed the request (13/10); the v1.3 JevBench submission is in the regular queue with no run published.

Fit / anti-fit

Fit when you need: a cheap hosted call that picks one of a few allowed actions and gives you a distribution, optionally from an image or a few video frames, for example in game loops, GUI agents or operational controls where your code already enumerates the legal moves.

Anti-fit when you need: more than eight options per question, inputs over 1,024 tokens, documented Noul or Score behaviour, local inference, or a model whose lineage and accuracy you can check independently.

Limits

Beta with no SLA and no third-party integrations.
Base model, size, architecture and training are undisclosed, so we can’t tell what the probabilities are calibrated against.
The 1,024-token input cap already fails long JevBench items, and rejected requests count as wrong in the vendor’s panel.
Noul and Score are listed by the vendor’s /v1/models but not demonstrated in the public docs.
The v1.3 panel includes tasks whose public training splits were in the training data, and every benchmark and latency number is vendor-run (UnverifiedClaim). Scores will be revisited when JevBench publishes a run.

Why it is in the catalog

MachineFi presents Trio-Spark as its own trained decision model, and it has a public, documented typed API that returns a probability for every option and, since v1.3, speaks the /v1/systemone shape. That meets the catalog criteria the same way hosted Jev and d1 do. Because lineage is undisclosed and the listing is vendor-submitted, it stays in with the caveats above and provisional scores. If it turns out to run another model underneath, it moves to runtimes.