OpenThai-SystemOne
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6.1/10
Apache-2.0 0.8B weights on HF with a pip package (openthai-systemone), a GitHub repo, an official Ollama build for /v1/systemone and about 14 quantized builds (GGUF, MLX, NVFP4, GPTQ, bnb). Very detailed card. Independent lab, no hosted API or SLA.
- Capability7.8/10
Choice (up to 255 options plus abstain), Score and Noul from a 256-slot head. Brier-loss calibration with per-type temperature and per-set ECE published. About 40 ms on an H100 and 23 ms per question through Ollama [vendor]. Benchmarks are author-run on public subsets [vendor].
- Adoption6/100
About 10.5k downloads and 33 likes on the main HF repo plus roughly 1k across quantized builds (6 Oct). Listed in the Ollama library as iapp/openthai-systemone. No production case study found.
Vendor claims
- Macro average 74.3 on the 13 public subsets from Bespoke Nimble's PUBLIC_BENCHMARKS (Nimble-9B 74.8, Jev 1.13.0 76.0 as published by Bespoke Labs on 18 Sep 2026); ahead of Nimble-9B on 8 of 13 subsets in the card's table[Vendor claim — not independently verified]
- Thai held-out sets: MASSIVE-th intent 90.0% (ECE 0.043), Prachathai67k topics 98.1%, XNLI-th 77.1%, Wongnai stars 63.5%; weakest sets Wisesight sentiment 51.6% and English banking77 45.4%[Vendor claim — not independently verified]
- Latency: about 40 ms for a 3-question Thai ticket on an H100, 154 ms on a MacBook M3 Max; through Ollama 0.35 the Q8_0 build runs 23 ms for one question and 136 ms for three[Vendor claim — not independently verified]
Read this first
- Every number here is author-run. iApp measured the public 13-subset bench with its own scripts and quotes Nimble-9B and Jev from Bespoke Labs’ published table. The model is not on JevBench or the Decision Index board yet, so treat all of it as UnverifiedClaim.
- Strong on Thai, uneven in English, and the card contradicts itself. The table bolds 8 of 13 subsets as ahead of Nimble-9B and shows SQuAD2 at 89.3, while the prose under it says the model leads on 4 of 13 and that SQuAD2 is at chance. It also flags BoolQ, PubMedQA and SummEval relevance (ECE 0.356 in the table, 0.79 in the text) as weak spots. We note the gap and do not pick a side.
- Ollama caps options at 26. The repo’s own API handles up to 255 options; the Ollama build is fine-tuned for Ollama’s prompt and drops wider questions.
What it is
OpenThai-SystemOne is an open Thai + English decision model from iApp Technology. It does not generate text. Given a state (text or JSON) and typed questions, it returns probabilities over the options in one forward pass, with a request and response contract that mirrors TypeSafe’s POST /v1/systemone.
| Type | You give | You get |
|---|---|---|
choice |
instructions + up to 255 options | choice, probabilities, confidence (slot 255 = abstain) |
score |
instructions + 2–10 ordered levels | fractional score, probabilities, confidence |
noul |
a yes/no question | p(yes) |
How it was built, per the card: the text tower of Qwen3.5-0.8B-Base (vision encoder removed), continued pretraining on about 5B Thai tokens, the LM head replaced by a 256-way slot head, training on about 2–3M decision examples, then a short calibration stage (Brier loss plus per-type temperature).
Specs
| Attribute | Value |
|---|---|
| Author | iApp Technology (Thailand) |
| Base | Qwen3.5-0.8B-Base text tower + Thai continued pretraining |
| License | Apache-2.0 |
| Weights | iapp/OpenThai-SystemOne + GGUF, MLX, NVFP4, FP8, GPTQ and bnb builds |
| Runtime | pip install openthai-systemone · Ollama ≥ 0.35 (ollama pull iapp/openthai-systemone) |
| Launch | (HF); Ollama build 1 Oct |
| Latency | 23–154 ms depending on setup — UnverifiedClaim |
| Pricing | Self-host (free weights) |
Numbers (UnverifiedClaim)
Public 13-subset bench, zero-shot, author-run:
| OpenThai 0.8B | Nimble-9B | Jev 1.13.0 | |
|---|---|---|---|
| Macro average | 74.3 | 74.8 | 76.0 |
Nimble and Jev figures are Bespoke Labs’ published numbers, not a same-day rerun. Through Ollama, iApp also ran Tev1 0.8B (64.3), Tev1 4B (74.9) and Nimble 9B (74.8) on the same records; on its 8 Thai sets the Q8_0 build scored 82.5 against 75.4 and 78.3 for the 4B and 9B.
Fit / anti-fit
Fit when you need: Thai or mixed Thai/English routing, moderation, intent and RAG relevance on a laptop or a small GPU, through code written for the TypeSafe SDKs or Ollama.
Anti-fit when you need: reading-comprehension yes/no, fine summary-relevance scoring, or a hosted API with an SLA.
Score working (6 Oct 2026)
- Maturity = availability 8 × 0.30 + docs 8 × 0.25 + integrations 6 × 0.25 + support 1 × 0.20 = 2.40 + 2.00 + 1.50 + 0.20 = 6.10
- Capability = types 10 × 0.30 + calibration 7 × 0.25 + latency 8 × 0.25
[vendor]+ accuracy 5 × 0.20 = 3.00 + 1.75 + 2.00 + 1.00 = 7.75 → 7.8 - Adoption = engagement 10 × 0.35 + ecosystem 8 × 0.35 + production 0 × 0.30 = 3.50 + 2.80 + 0 = 6.3 → 6
Calibration 7: method (Brier loss + per-type temperature) and per-set ECE are published. Accuracy 5: author-run on public data, not an independent board.