OpenThai-SystemOne-Vision GGUF (v1.5)
Developed by iApp, launched with AIEAT. Part of OpenThai (formerly OpenThaiGPT).
OpenThai-SystemOne-Vision v1.5 for stock llama.cpp: the open Thai +
English System One decision model (0.86B parameters) that answers typed questions about a state, with or without a
screenshot, in one forward pass and without generating text. llama-server serves it through its /v1/systemone endpoint.
| file | size (bytes) | sha256 |
|---|---|---|
OpenThai-SystemOne-Vision-Q8_0-decision.gguf |
1,320,471,008 | 3ee8ac840ae1421afbde362f6d869008ce37571a281c9edcc3578b44bfec3806 |
mmproj-OpenThai-SystemOne-Vision-F16.gguf |
204,986,656 | 7d25d180974d7a570bc1ef49afbd984da265fcb1f54704c2d20c20b569db3c6e |
The first file is the text tower at Q8_0 with the decision readout (one row per option label, taken from the model's slot head, plus the per-type temperatures). The second is the vision tower and merger at F16; it is only needed when a request carries an image.
Both files change with every version (the vision tower is trained too): always take the two from the same commit.
| version | tag | commit |
|---|---|---|
| v1.5 (2026-11-20) | v1.5 |
the commit the tag points to |
| v1.4 (2026-11-15) | v1.4 |
0f5e1ffae67c86b59eedc3433d23261714937d70 (Q8_0 file 1,320,471,072 bytes, sha256 5b022149…59a0; mmproj sha256 272385cd…c817) |
Later checkpoints are added as new commits. Pin the commit you tested if your app must not change behind its users' backs.
Run
Tested with llama.cpp build b12242.
llama-server -m OpenThai-SystemOne-Vision-Q8_0-decision.gguf \
--mmproj mmproj-OpenThai-SystemOne-Vision-F16.gguf -c 32768 --parallel 4 -ngl 99
curl -s http://localhost:8080/v1/systemone -H 'content-type: application/json' -d '{
"state": "os: windows\napp: CalculatorApp.exe\ngoal: กดปุ่มบวก",
"questions": {"target": {"type": "choice",
"instructions": "Which element should be acted on to achieve the goal?",
"criteria": {"0": "button \"Seven\"", "1": "button \"Plus\"", "2": "button \"Equals\"", "none": "none of these"}}}
}'
The answer holds the chosen option and a probability for every option. Question types: choice (up to 255 options),
noul (yes / no; only instructions) and score (criteria is a list of 2 to 10 level names). The prompt layout is inside
the GGUF, so the client sends the request as it is.
Clicking with it (computer use)
This build has no point head: llama-server does not expose what that head reads. A desktop app clicks by asking a choice
question over the controls it reads from the screen through the accessibility API, as in the example above:
- state:
os: windows | macos | linux,app: <program>,goal: <the command>, one per line; - options: the control's type in lower case and its name in quotes (
button "Save",menuitem "View",listitem "General"), in reading order, with"none": "none of these"last; - no screenshot is needed for this question, and Thai commands are answered slightly better without one.
What to expect, measured with the PyTorch weights on our own test screens (lists of up to 120 controls, no screenshot; the test controls' names are never asked in training); this build picked the same option as the PyTorch weights in 359 of 360 checks:
| the command | v1.4 | v1.5 |
|---|---|---|
| Windows 11 (about 1,650 targets) | ||
· names the control as written on screen, in English or after a Thai verb (กดปุ่ม Save) |
99.0 | 99.5 |
| · a natural English sentence | 96.0 | 98.0 |
| · a natural Thai sentence, English labels on screen | 86.1 | 93.2 |
· control not listed: answers none (natural Thai / natural English / name as written) |
8 / 9 / 15 | 51 / 67 / 75 |
Ubuntu 24.04, GNOME apps (os: linux, 178 targets) |
||
| · names the control as written | 98.9 | 99.4 |
| · a natural Thai sentence | 78.8 | 89.4 |
| · a natural English sentence | 89.4 | 90.6 |
· control not listed: answers none (natural Thai / natural English / name as written) |
10 / 11 / 11 | 58 / 79 / 84 |
On macOS control lists (560 targets, os: macos) v1.4 chose the control named in the command 99 % of the time; v1.5 was
trained on those screens and has no held-out macOS figure yet.
When the thing asked for is not on the screen, v1.5 answers none for about half to three quarters of such commands
(v1.4: 8 to 15 %) and otherwise still picks some control. Keep a check in the app before it clicks, for example that the
chosen control's name is related to the command, and ask the user when it is not. The answer carries a probability for every
option, so an app can also treat none with a plausible second choice as "ask about that control".
The same model reads voice commands (see the main card): "click" + a name on the screen is now understood as a click 95.2 % of the time (v1.4 78.1 %).
License
Apache-2.0, as the source model, built on Qwen3.5-0.8B-Base. Developed by iApp, launched with AIEAT.