Thai language, legal, decision, document and speech models, compared task by task. Within each card, every model was run under one protocol on the same test data, so scores compare only inside a card. Every number links to its source, and the ranking, averages and "best" marks are computed by this page.
Who runs this. OpenThai builds some of the models listed here (tagged OpenThai). We ran the competing models ourselves with the same scripts and settings, we keep the results where they beat us, and each card says what it does not measure. Treat this as a well-documented self-evaluation, not an independent audit; the sources and harnesses are linked so anyone can check or re-run them.
Loading…
▲ ahead · = tie · ▼ behind, counted test set by test set. Hover a row for what it counts; the card has the detail.
Loading results…
Only numbers inside one card. A card is one protocol: the same test data, prompts or decoding settings, scorer and normalisation for every model in it. Different cards use different data and metrics, so do not read across them, even when two cards share a metric name such as CER.
The team that built the OpenThai model in each card ran every model in that card, unless the card says otherwise (the 2025 archive is one published run from a technical report). Hosted APIs are called on their public endpoints with default settings; open models are run from their published checkpoints. Each card's "How this was measured" lists the exact settings.
Where a card shows an interval, it is the 95% confidence interval that the evaluation reported. Head-to-head rows give the difference between two models and, where measured, its interval. "Tie" means the gap is within measured noise: the interval includes zero, or repeat runs vary by more than the gap.
Open an issue or pull request on the OpenThai eval harness, write to us through the contact page, or ask in Discord. We add models by running them ourselves under the card's protocol; self-reported scores are not copied in.
OpenThai