OpenThai-SystemOne is a small specialist I want on the shelf next to chat models and tool-call glue: an 0.8B decision and routing SLM that answers typed questions with calibrated probabilities instead of free-form prose.
iApp Technology and OpenThai built it on Qwen3.5-0.8B-Base. They continued Thai and English pretraining on the text tower, then replaced the usual language-model head with a 256-way slot softmax. You hand the model a state (plain text or JSON) plus one or more typed questions. In a single forward pass you get a distribution over the options you declared. The license is Apache-2.0, and the weights sit on Hugging Face for anyone to pull.
What the model actually does
The contract is narrow on purpose. Three question types cover most of the agent-routing work I care about:
- choice. Instructions plus up to 255 named options. You get the selected option, the full probability vector, and a confidence score.
- score. Instructions plus two to ten ordered level descriptions. You get a probability-weighted fractional score, plus probabilities and confidence.
- noul. A yes/no question. You get p(yes).
There is no free-form generation. Context is text-only, up to 64K tokens per request, in Thai and English. Slot 255 is abstain when none of the options fit. The request shape mirrors TypeSafe’s System One API, so code written for that SDK can point at this checkpoint with little ceremony.
That design matches the jobs people actually ship: ticket triage, intent detection, moderation flags, RAG relevance, verifying another model’s output, and picking the next tool or UI action for a computer-use agent. You declare the menu. The model ranks the menu. You keep the policy of what to do with a low-confidence answer.
A fast three-day climb (vendor numbers)
Initial weights landed September 20 as v0.1. v0.2 followed on the 21st. v0.3 on the 22nd is the material quality update. On the vendor’s public 13-subset macro, they report 61.9 → 63.2 → 74.3 across those three days. Treat those as vendor-reported until you reproduce them; the jump into the mid-70s is still the signal that continued SFT and recalibration lifted this specialty.
Latency claims are also vendor-reported: about 40 ms for a short Thai ticket on an H100, and 154 ms on a MacBook M3 Max over MPS. The model card also notes roughly 40 to 48 ms batch-1 latency with 255 options on a shared H100. Measure on your own hardware before you promise a product budget.
Where it would sit on any Small AI shelf
A useful Small AI shelf still has three drawers that matter for agents: a MiniCPM5-2B-class model in the roughly 2 to 4B generalist drawer for local chat and broader agent work, a Needle 3-class tens-of-MB specialist for on-device tool glue, and a typed decision head for routing and verification. OpenThai-SystemOne fills that third slot at 0.8B, with Thai strength baked into the continued-pretrain mix.
I do not build products for Thai today. If I did, this would be a must-have choice for that routing tier. Even without a Thai product line, the pattern is still useful anywhere you need choice, score, and noul with probabilities you can threshold.
If you are wiring an agent, that split earns its keep. Let a generalist draft or reason when you need language. Let a tiny tool model emit calls when the catalogue is local and small. Let a decision head like this one pick the department, the next tool, or whether a passage actually answers the question, with probabilities you can threshold.
A nearby watch-point, not the story
On September 22 Alibaba also talked up Qwen Intelligence, a full-stack agentic phone solution, with HONOR as first partner ahead of Magic9 / Robot Phone coverage around September 28. Open mobile foundation sizes are not disclosed yet. Until those weights and sizes show up in public, I read Qwen Intelligence as a distribution and product story. OpenThai-SystemOne is a different artifact: a calibrated decision head you can download and point at agent routing today.
Limits worth writing down
The card is honest about weak spots. Summary-relevance scoring and some fine-grained English intent sets still lag. English calibration is softer than Thai. The model will not do multi-step reasoning or arithmetic for you. Cap options at 255 per question (bucket into stages if you need more), stay text-only, and route low-confidence cases to a larger model or a human. Those are product constraints, not surprises.
Time to dig deeper
Start with the OpenThai-SystemOne card on Hugging Face, skim the version table for v0.1 through v0.3, and note the Apache-2.0 license and the TypeSafe-compatible /v1/systemone shape.
Then do the useful thing. pip install openthai-systemone, load the checkpoint, and feed it a real support ticket or agent state from work you already care about. Ask one choice, one score, and one noul in the same call. Log the probabilities and the confidence. Flip the option order a few times and see whether your client’s order-invariant mode matters for your taxonomy. Time a short request on whatever GPU or Apple Silicon you have. Compare a low-confidence routing decision against a MiniCPM5-2B-class generalist and a Needle-sized tool specialist on the same state. That afternoon will teach you more about this typed-decision tier than another launch macro.