Skip to content
TIAPartner
AccountStart freeVI

Decision model family

TIA 1.5

Recorded on

An experimental TIA build evaluated locally on 7 October 2026, scoring 125/231 on the public JevBench items. It has not passed the promotion gate and is not served by the API. This is not a result or ranking certified by JevBench.

Try the currently served build in PlaygroundCreate a free account
TIA 1.5 at a glance

This build has no row on the Benchmark table yet. The measurements, splits and dates are in the release notes below.

Version notes: TIA 1.5

  • Local evaluation on the 231 public JevBench items: 125/231 (54.1%); easy 44/48, original 44/72, hard 37/111. Comparisons use published results from other systems, not a fresh evaluation of every model.
  • Standard Vietnamese decisions (vi-dev): reaches 82.3% (vs 66.6% for TIA 1.4 mini), Vietnamese legal questions (vn-law) reaches 96.9%, and scam detection (vi-sms) reaches 97.4%.
  • Internal jevlike-dev recorded an argmax accuracy increase from 0.3111 to 0.3652. The experimental build nevertheless failed the promotion gate's improvement criterion; this accuracy change does not override the gate.
  • Choice accuracy: td-dev reaches 0.765 vs 0.747, gen-dev reaches 0.906. Calibrated Brier score: 0.0680 (lower is better).
  • Recorded local GPU p50 is 37 ms on an RTX 5080. This is not API latency or a Windows CPU result.

Try it in the PlaygroundSee the benchmark tableRead the API docsEnterprise deployment

Builds

TIA 1.5

Role
Experimental build evaluated on 7 October 2026, not promoted to the API. Its local measurements do not describe the served model.
Accuracy
not yet on the public bench
KL
not yet on the public bench

We have not measured this build on the public bench, so we print no numbers. When we do, they land on the Benchmark page with their split and date.

TIA 1.4 mini

Role
Predecessor build (1 Oct 2026), reached 102/231 on JevBench.
Accuracy
not yet on the public bench
KL
not yet on the public bench

We have not measured this build on the public bench, so we print no numbers. When we do, they land on the Benchmark page with their split and date.

Need TIA inside your own infrastructure? That is arranged case by case. Contact us through the Partner or invest page

Latency

p50 0.88 s, p95 1.16 s, end to end from one PC in Vietnam to the public demo endpoint: 30 sequential calls, 5 questions each, measured on Oct 1, 2026. It includes the network. One location and one connection, so your numbers will differ; this is not a guarantee.

Versions and experiments

  1. TIA 1.5 (TIA 1.5, TIA 1.4 mini)

    An experimental TIA build evaluated locally on 7 October 2026, scoring 125/231 on the public JevBench items. It has not passed the promotion gate and is not served by the API. This is not a result or ranking certified by JevBench.

  2. TIA 1.4 mini

    Predecessor build that served the System One API from Oct 1, 2026 until Oct 3, 2026, before being succeeded by TIA 1.5.

  3. TIA 1.3 mini

    A new mini that replaces TIA 1.2 mini behind the System One API. It spots real Vietnamese scam SMS, follows stated rules better and depends less on option order; the call and the model id stay the same.

  4. TIA 1.2 mini

    A new mini that replaces TIA 1.1 mini behind the System One API. Clearly better Vietnamese on TIA's held-out set; the call and the model id stay the same.

  5. TIA 1.1 mini

    A new mini that replaces TIA 1.0 mini behind the System One API. The call and the model id stay the same.

  6. TIA 1.0 (TIA 1.0, TIA 1.0 mini)

    The first release of the TIA family. Send one state with several typed questions (yes/no, pick one, score) and get a calibrated probability distribution for each of them, in one call.

This list includes historical builds and experimental candidates. TIA 1.5 is not approved for API deployment; the API currently serves revision eff27e8, with a historical JevBench result of 122/231. A recorded date or local measurement does not establish a release.

How we version

  • Names are the family plus a major.minor number: TIA 1.5. Minor goes up for improvements that keep the API call the same; major goes up when behaviour or the contract changes.
  • A build name follows the version number, for example “mini”. A build only gets numbers once it has been measured.
  • Every release ships notes: what changed, which split it was measured on, and when.
  • Every number names its split. Today that is the internal dev set; the official test split has not been read, and dev numbers are not comparable with test numbers.

Send TIA your first state

The Playground runs the real model with no account. When you need an API key, create a free account.

Try the currently served build in PlaygroundCreate a free account