OpenThai-SystemOne-Vision (v1.4)

Developed by iApp, launched with AIEAT. Part of OpenThai (formerly OpenThaiGPT).

The vision variant of OpenThai-SystemOne: the same open Thai + English System One decision model, now with images in the state and a fourth question type, point, that returns where on a screenshot something is. Screenshot in, click out, in one forward pass, no text generation.

v1.4 (2026-09-28) clicks better on Windows apps (Settings, Calculator) and on icons, and understands voice commands a little better; it plays Doom from pixels worse than v1.3 (see Evaluation). v1.3 (2026-09-26) clicked better on real Windows and Mac apps (trained on live Windows 11 and macOS screens) and on large screens, and adds point_refine; v1.1 (2026-09-24) added Thai voice-command understanding for hands-free computer control (a spoken Thai command + the app and window title in, the operation and its target out) and stronger desktop and Doom play. See Versions.

	OpenThai-SystemOne (text)	OpenThai-SystemOne-Vision
parameters	0.75B	0.86B (ViT 100M + point head 9M)
question types	choice / score / noul	choice / score / noul / point
state	text or JSON	text or JSON + up to 4 images
typical use	routing, moderation, RAG relevance, tool choice, agents with element tables	GUI agents from pixels, Thai voice control, Thai screen reading, game frames, documents
text-only behaviour	v0.3	same API; text scores equal or better than v0.3 (tables below)

API (superset of the TypeSafe /v1/systemone contract)

{
  "state": "เป้าหมาย: ค้นหาเที่ยวบิน BKK→CNX",
  "images": [{"id": "screen", "data": "data:image/png;base64,...", "role": "screenshot"}],
  "questions": {
    "where":  {"type": "point",  "instructions": "ปุ่มค้นหาเที่ยวบิน", "image": "screen"},
    "op":     {"type": "choice", "instructions": "next operation", "criteria": {"CLICK": null, "TYPE_TEXT": null, "SCROLL_DOWN": null, "DONE": null}},
    "loaded": {"type": "noul",   "instructions": "is the page fully loaded?"}
  }
}

point answers: {"x": 0.63, "y": 0.41, "confidence": 0.86, "abstain": 0.02, "candidates": [{"x", "y", "p", "bbox"}, ...]} (normalised image coordinates; abstain = probability that the target is not on the image; candidates = the above-threshold token clusters of the attention heat-map, GUI-Actor style).

Limits: ≤ 4 images, ≤ 8k visual tokens per request (a 1280×800 screenshot ≈ 1,000 tokens at 32 px/token; a 320×240 game frame ≈ 80); "detail": "high" allows up to 1920×1200.

Usage

Code status (2026-11-15). The Python client shown on this card (image inputs, point, point_refine) is not in the public GitHub repository yet; it arrives with the next code update there. What runs today with public code: the GGUF build iapp/OpenThai-SystemOne-Vision-GGUF on stock llama.cpp (llama-server, /v1/systemone): choice / score / noul questions, with or without a screenshot. It has no point head, so a click there is a choice among the controls the app reads from the screen (see that card).

from openthai_systemone import SystemOneClient
c = SystemOneClient("iapp/OpenThai-SystemOne-Vision")
r = c.system_one("goal: open the login form", {"where": {"type": "point", "instructions": "ปุ่มเข้าสู่ระบบ", "image": "s"}},
                 images=[{"id": "s", "path": "screen.png", "role": "screenshot"}])
print(r.answers["where"].x, r.answers["where"].y, r.answers["where"].confidence)

Thai voice control

The voice-command schema (21 operations such as SEARCH, CLICK_NTH, OPEN_WEBSITE, QUIT_APP, PLAY_PAUSE, LIKE, COMMENT, NONE = not a command, plus search_site / app / site / key / nth slots) and a macOS controller (on-device Thai speech recognition → this model → native mouse / keyboard events) will be published in the GitHub repo under demo/voice_control/. One text-only call decides the operation; clicks add one screenshot call.

state = 'คำสั่งเสียงภาษาไทยจากผู้ใช้ที่ควบคุมคอมพิวเตอร์ด้วยเสียง: "อยากดูเนเน่ร้องเพลง"\nโปรแกรมที่ใช้อยู่: Google Chrome\nหน้าต่าง: Facebook'
# questions = intent_schema.intent_questions(apps)  -> operation SEARCH, search_site YouTube

Evaluation

v1.4 next to v1.3, same scripts; single option order unless noted; accuracy in %. "default" = the plain API call (screenshot at ≤ 1280×800); "+ refine" = screenshot at 2 MP ("detail": "high") with point_refine over the top-3 first-pass candidates, landscape screens only (zoom 50 % on ScreenSpot-v2 / Thai GUI / OS-Atlas, 35 % on Windows / Mac live and ScreenSpot-Pro). On desktops and large screens, send "detail": "high" and point_refine.

Live desktops: real apps, click targets from the accessibility tree

benchmark	v1.3 default	v1.3 + refine	v1.4 default	v1.4 + refine	Qwen3-VL-2B	Qwen3-VL-4B
Windows live (319 targets, 28 screens of Windows 11: Settings, File Explorer, Control Panel, Notepad, Paint, Calculator, Store, Edge, …)	68.3	69.3	70.8	70.5	67.1	73.0
· Windows Settings pages (114)	69.3	–	74.6	–	66.7	68.4
· Windows Calculator (20)†	60	–	80	–	80	95
Mac live (358 targets, 29 screens: System Settings, Finder, Safari, Office, iWork, …)	63.7	67.6	64.5	67.3	67.0	67.9
· label-checked (340: 18 targets found unanswerable removed)	66.8	71.2	67.6	70.9	70.6	71.5
latency per click (H100, eval harness)	~70 ms	0.1–0.25 s	~70 ms	0.1–0.25 s	~0.9 s	~0.75 s

The Windows and Mac live screens are never trained on; the training data includes other screens of the same apps. † v1.4's synthetic calculators use Windows' standard accessible names ("Two", "Multiply by"), which is how the Windows live Calculator targets are named, so part of that gain is learning the naming convention.

GUI grounding benchmarks (point: predicted click inside the target box)

benchmark	v1.3 default	v1.3 + refine	v1.4 default	v1.4 + refine	Qwen3-VL-2B	Qwen3-VL-4B
ScreenSpot-v2, all (1,272)	82.2	85.2	82.8	85.9	85.8	90.4
· desktop text / icon	90.2 / 75.7	91.8 / 83.6	89.2 / 77.1	91.2 / 85.0	93.8 / 82.9	96.9 / 85.7
· mobile text / icon	94.5 / 71.6	93.8 / 74.9	94.8 / 71.6	94.5 / 75.4	94.1 / 82.0	96.9 / 83.4
· web text / icon	84.2 / 70.0	88.5 / 74.9	85.9 / 71.4	89.3 / 76.4	85.9 / 72.4	91.9 / 83.7
ScreenSpot-Pro (1,581, 26 pro apps up to 4K; Windows / macOS / Linux)	9.0	32.9 (29.3 / 38.7 / 28.0)	8.7	32.8 (29.2 / 38.9 / 26.0)	43.4*	54.8*
Desktop (OS-Atlas macOS / Windows / Linux, held-out)	53.7	59.8 (600)	53.8	59.3 (600)	63.8	65.6
Desktop "not on screen" abstention (26)	92.3	–	92.3	–	–	–
Thai GUI held-out domains, point (1,611)	77.8	76.7 (600)	78.6	74.9 (600)	72.4	80.4
Thai GUI set-of-marks choice / operation	93.1 / 99.6	–	92.6 / 99.8	–	–	–
Thai screen QA noul / choice	93.1 / 85.9	–	94.7 / 85.0	–	–	–
Thai GUI "not on screen" abstention (179)	48.0	–	49.7	–	–	–

Qwen3-VL numbers: our runs with the same harness (screenshot + instruction, click inside the box); * every 2nd ScreenSpot-Pro item (790). Mac live and the 600-item Thai GUI / OS-Atlas subsets move by ±1.5 points between settings; smaller differences are noise.

Better clicks on large screens: point_refine

r = c.system_one("goal: …", {"where": {"type": "point", "instructions": "ปุ่มบันทึก", "image": "s"}},
                 images=[{"id": "s", "path": "screen.png", "role": "screenshot", "detail": "high"}],
                 point_refine={"zoom": 0.35, "cands": 3})          # HTTP API: "point_refine": {...} or true

The first pass points on the whole screen; the client then re-points on a crop around each of the top candidates at native detail and keeps the most confident. It only zooms landscape screens larger than the pixel budget (zooming phone screenshots hurts). v1.4: ScreenSpot-Pro 8.7 → 32.8, ScreenSpot-v2 82.8 → 85.9, Mac live 67.6 → 70.9 (checked). Windows windows usually fit the 2 MP budget, so there it mostly adds the higher resolution. Cost: 1 + cands forward passes.

Thai voice commands (operation + slots from one spoken sentence)

set	v1.1	v1.3	v1.4
real commands spoken in use, operation (183)	88.5	89.6	90.7
· slots (site / app / key / Nth) when the operation is right	100.0	100.0	100.0
· chit-chat between commands acted on (41), lower is better	7.3	7.3	7.3
real off-command conversation picked up by the microphone, acted on (227), lower is better	2.6	4.4	4.0
synthetic voice commands, held out (1,406 questions)	90.5	92.0	91.8
synthetic hard cases (fragments, site vs app, ordinals, 446 questions)	90.8	93.3	92.6

Doom from pixels (demo/doom_demo.py, defend_the_center, skill 3, two 320×240 frames, 40 episodes per seed)

seed	v1.1	v1.3	v1.4	ms / decision
999	11.8	13.0	11.4	~50
777	–	11.9	9.7	~50
mean of both (80 episodes)	–	12.5	10.6
the privileged scripted expert it imitates	6.8	6.8	6.8	–

Known regression: v1.4 plays Doom from pixels worse than v1.3 (−1.9 kills per episode). Use v1.3 (revision="v1.3") for the Doom demo. Offline Doom decisions (600 frames), v1.4: move 88.5, attack 97.2, in-danger 96.7, nearest-enemy point 48.9, enemy-absent 98.3.

Text-only tasks

Public 13-subset benchmark (the Bespoke / Jev public suite; noul/choice = accuracy, score = exact level):

subset	v0.3	v1.3	v1.4	Jev 1.13.0*	Nimble-9B*
aegis2	83.2	82.0	81.2	80.4	81.2
boolq	79.7	79.7	79.0	89.7	86.0
civil_comments	79.0	86.3	86.3	81.0	70.3
helpsteer2 (score)	41.6	40.4	40.4	34.1	39.0
massive-de-DE	88.3	89.7	90.0	86.9	83.4
massive-en-US	88.3	90.3	90.3	87.4	86.9
multinli	89.0	88.3	89.0	82.9	85.3
paws	94.0	94.0	94.0	89.2	82.8
pubmedqa	64.0	67.2	67.2	77.2	75.6
squad2	89.3	88.6	88.3	82.9	80.6
summeval-consistency (score)	75.0	75.0	75.7	81.2	75.7
summeval-relevance (score)	21.7	20.4	20.4	35.0	49.2
vitaminc-dev	72.5	71.0	70.8	80.1	76.6
macro, 13 subsets	74.3	74.8	74.8	76.0	74.8
macro, 10 accuracy subsets	82.7	83.7	83.6	83.8	80.9

* published by Bespoke Labs for their own harness on the same subsets (Jev via the TypeSafe API).

Thai held-out and eval-split sets (th unless noted; banking77 and xLAM are English):

set	v0.3	v1.3	v1.4	v1.4, 8 orders
banking77 (77-way, whole dataset held out)	45.4	62.0	62.1	64.4
contrastive pairs, one-fact flips	80.7	82.1	81.8	80.1
MASSIVE-th intent	90.0	91.1	91.2	92.0
Prachathai topics	98.1	98.6	98.6	98.7
SIB-200 Thai topic (whole dataset held out)	77.9	75.5	75.0	77.0
Wisesight sentiment (whole dataset held out)	51.6	52.5	52.2	51.7
Wongnai stars (score, exact)	63.5	63.3	62.9	62.9
xLAM tool selection	99.4	99.4	99.4	99.3
XNLI-th	77.1	75.5	75.6	75.7
macro, 8 accuracy sets	77.5	79.6	79.5	79.9

Latency (batch 1, H100, end-to-end SystemOneClient)

request	tokens	ms
two 320×240 game frames + short state, 2 questions	298 (160 visual)	50
1280×800 screenshot + goal, point + choice	1,062 (1,000 visual)	52
large screenshot at "detail": "high" + point_refine (3 candidates), eval harness	~2,000 visual per pass	~135 (FHD) – 245 (4K)
text only (3-question Thai ticket)	166	~40

Versions

version	date	change
v1.4	2026-09-28	+ icon grounding and "not on screen" refusals from xlangai/Jedi (20k icon snaps, 20k refusal rows), 12k synthetic calculators / title bars / toolbars / media controls labelled with their standard accessible names (English + Thai), Mac and Thai GUI rehearsal up; from v1.3. Windows live 68.3 → 70.8 (Settings 69.3 → 74.6, Calculator 60 → 80), ScreenSpot-v2 82.2 → 82.8 (85.9 with refine), Thai not-on-screen 48.0 → 49.7, voice 89.6 → 90.7; Doom from pixels 12.5 → 10.6 kills; text, Thai, Mac live, ScreenSpot-Pro, OS-Atlas unchanged within noise.
v1.3	2026-09-26	+ live desktop grounding: 665 records from 319 live Windows 11 screens (UI Automation targets; Settings, Control Panel, dialogs, shell folders, Calculator modes, Store, Edge) and 386 records from 116 live macOS screens (accessibility targets; System Settings, Finder, utilities, Safari, Word / Excel / PowerPoint / Pages ribbons), with account names blacked out; on top of an internal v1.2 round (ShowUI desktop + web, 40k zoomed crops, 75k xlangai/Jedi records, 2 MP training, averaged with v1.1). Client / API point_refine. Windows live 63.3 → 68.3, ScreenSpot-v2 80.7 → 82.2 (85.2 with refine), ScreenSpot-Pro ~10 → 32.9 with refine, OS-Atlas desktop 50.6 → 53.7, Doom 11.8 → 13.0, text and Thai unchanged.
v1.1	2026-09-24	+ Thai voice-command intents (28k synthetic training records from 21k distinct commands + 8.9k hard-case records: fragments / chit-chat = NONE, site vs app, ordinal + topic, ASR spelling variants), desktop grounding round (OS-Atlas macOS / Windows / Linux), text-anchor round (civil_comments, VitaminC); temperatures recalibrated. Voice 65 → 88.5, desktop 39 → 51, Doom pixels 5.5 → 11.9 kills, ScreenSpot-v2 82.7 → 80.7.
v1.0	2026-09-23	first vision release (Thai GUI crawl, public grounding, Doom rollouts, screen QA on the v0.3 text tower)

How it was built

1. Text tower + slot head from OpenThai-SystemOne v0.3 (Qwen3.5-0.8B, Thai CPT 4.5B tokens, decision SFT + calibration).
2. Vision tower + merger from Qwen3.5-0.8B-Base, re-aligned with an LM loss on 200k+ image→text pairs (Thai rendered OCR pages in 55 fonts, Thai web screenshots → element tables, Doom frames → object tables, public screen/document QA).
3. Decision SFT on the v0.3 text mix plus Thai GUI grounding (Playwright crawl of 300+ Thai sites, teacher-written and blind-verified Thai/English instructions), public English grounding (Wave-UI, GroundUI, ShowUI), desktop grounding (OS-Atlas), Doom rollouts (privileged expert, 1.5M frames) with move / attack / in_danger / nearest_enemy questions, and screen QA.
4. v1.1: a text-anchor round and a desktop round averaged into one checkpoint, then a voice-command round (label-conditioned generation by a teacher LLM, every sentence blind re-labelled and kept only on agreement) and a hard-case round aimed at the mistakes seen in real use; the rest of the mix is rehearsed at the same time.
5. v1.2 (internal): three grounding rounds from v1.1 (ShowUI desktop + web and 2 MP screenshots; 40k zoomed crops at native detail that teach the point_refine second pass; 75k xlangai/Jedi records), with voice, text and Doom rehearsed; the release is the equal-weight average of v1.1 and the last round, which kept v1.1's Thai GUI and Doom skills and the new desktop gains.
6. v1.3: one round from v1.2 with live Windows 11 and macOS screens captured through the accessibility APIs (a small agent inside a Windows VM; the macOS accessibility tree), other apps / pages than the live eval screens, personal names blacked out, rest of the mix rehearsed.
7. v1.4: one round from v1.3 with icon grounding / refusal data from xlangai/Jedi and synthetic symbol screens, with the Mac and Thai GUI data weighted up so they do not slip.
8. Calibration of the per-type temperatures.

License & credits

Apache-2.0, built on Qwen3.5-0.8B-Base. Developed by iApp, launched with AIEAT. GUI grounding recipe after GUI-Actor; System One contract after TypeSafe AI's Jev. Grounding data from OS-Atlas, ShowUI, Wave-UI, GroundUI and Jedi (OSWorld-G); benchmarks ScreenSpot-v2 and ScreenSpot-Pro.