Key enhancements over Typhoon OCR 1.5
- Layout with bounding boxes. Each element comes with a category (13 types, such as title, text, table, formula, header and footer) and a box, in reading order. Typhoon OCR 1.5 returned the page as Markdown without positions.
- General visual question answering. The model answers free-form questions about a document and also handles document classification, key-information extraction to JSON and chart-to-data tasks. On ThaiOCRBench, which covers these tasks, the score rises from 40.5 to 59.9.
- Higher accuracy on public Thai and English benchmarks: MDPBench from 62.8 to 79.8 (Thai pages from 64.3 to 81.3), olmOCR-bench from 62.0 to 84.7, OmniDocBench v1.6 from 70.97 to 90.97, and SEA-DocBench NED from 0.1726 to 0.1153 (lower is better).
- Better on photographed pages: MDPBench photographed pages from 61.6 to 79.7.
- Markdown is still available. A few lines of code convert the JSON to Markdown, with or without headers and footers (see Usage).
- A smaller 0.8B model for page parsing at lower cost.
Model details
Table | |
|---|
| Parameters | 2B, BF16 |
| Architecture | Qwen3.5 (Qwen3_5ForConditionalGeneration) |
| Base model | infly/Infinity-Parser2-Flash |
| Training | Supervised fine-tuning (13,500 steps), then GRPO with a LoRA adapter (r=64, 500 steps), merged into the weights in fp32 and stored in bf16 |
| Languages | Thai and English. The training data also contains Chinese, which we did not evaluate separately. |
| Input | One page image. Trained with images of up to 4 MP; the benchmark numbers below use up to 16.7 MP. |
| Output | A JSON list of layout elements in a json code block |
Comparison with the base model
Both models were run with the same prompt, settings and scripts, with vLLM 0.27.1 in eager mode.
Table with columns: MDPBench (higher is better), Infinity-Parser2-Flash, Typhoon OCR V2 2B, Change| MDPBench (higher is better) | Infinity-Parser2-Flash | Typhoon OCR V2 2B | Change |
|---|
| Thai pages | 61.3 | 81.3 | +20.0 |
| English pages | 75.4 | 78.4 | +3.0 |
| Photographed pages | 65.3 | 79.7 | +14.4 |
| Digital pages | 78.9 | 81.3 | +2.4 |
| Overall | 68.3 | 79.8 | +11.5 |
We are re-running the base model on the other four benchmarks in eager mode, and we will add them to this table.
Results
We ran every model below ourselves with the same evaluation scripts. The numbers are not copied from other papers or model cards, so some differ from what the other authors report.
- Typhoon OCR V2 and Infinity-Parser2-Flash use the doc2json prompt shown in Usage, images of up to 16.7 MP, greedy decoding, at most 16,384 new tokens and vLLM 0.27.1 in eager mode (for Infinity-Parser2-Flash, see the note under the table). Their JSON is converted to Markdown with Infinity-Parser2's conversion code, which drops headers, footers and page footnotes. The model still outputs these elements; only the conversion leaves them out. Usage shows how to keep them.
- The other models use the prompts and post-processing published with them.
- MDPBench and SEA-DocBench are limited to their English and Thai subsets.
- PaddleOCR-VL-1.6 does not support VQA, so it has no ThaiOCRBench score.
In each column, bold is the best result and underlined the second best. ↑ means higher is better and ↓ lower is better.
Table with columns: Model, Params, MDPBench ↑, MDPBench TH ↑, olmOCR-bench ↑, ThaiOCRBench ↑, SEA-DocBench NED ↓, OmniDocBench v1.6 ↑| Model | Params | MDPBench ↑ | MDPBench TH ↑ | olmOCR-bench ↑ | ThaiOCRBench ↑ | SEA-DocBench NED ↓ | OmniDocBench v1.6 ↑ |
|---|
| Chandra-OCR-2 | 5B | 74.2 | 71.1 | 84.4 | 26.4 | 0.1141 | 89.65 |
| dots.mocr | 3B | 76.2 | 73.6 | 85.2 | 37.9 | 0.1135 | 91.09 |
| PaddleOCR-VL-1.6 | 0.9B | 72.5 | 68.4 | 80.8 | n/a | 0.1323 | 95.83 |
| Infinity-Parser2-Flash† | 2B | 68.3 | 61.3 | 84.4 | 35.2 | 0.1906 | 86.44 |
| Typhoon OCR 1.5 | 2B | 62.8 | 64.3 | 62.0 | 40.5 | 0.1726 | 70.97 |
| Typhoon OCR V2 0.8B | 0.8B | 77.4 | 77.2 | 81.7 | 36.4 | 0.1081 | 89.37 |
| Typhoon OCR V2 2B | 2B | 79.8 | 81.3 | 84.7 | 59.9 | 0.1153 | 90.97 |
† Infinity-Parser2-Flash shares our models' architecture, so we re-ran it in eager mode. Its MDPBench score is from eager mode (68.3; it scored 61.5 in vLLM's default compiled mode). Its other four scores are still from compiled mode and are probably too low. Eager re-runs are in progress, and we will update them here.
The 2B model has the highest MDPBench, MDPBench Thai and ThaiOCRBench scores of the models we evaluated. On English parsing it is close to the strongest open parsers but does not lead: dots.mocr is ahead on olmOCR-bench (85.2 against 84.7), and PaddleOCR-VL-1.6 is ahead on OmniDocBench (95.83 against 90.97), mainly because of tables and formulas. On SEA-DocBench the smaller Typhoon OCR V2 0.8B has the lowest NED (0.1081 against 0.1153).
Per-benchmark component scores will be in the technical report (coming soon).
Examples
All six examples are outputs of this model on benchmark items, run with vLLM 0.27.1 in eager mode. The two parsing examples use max_pixels at 4 MP, the image size the model was trained with; the other four come from our ThaiOCRBench run, where each item's question is the prompt.
1. Full-page parsing: a dense Thai exam page
Left: the input (MDPBench). Right: the predicted boxes, numbered in reading order. 19 of the 29 predicted boxes match the benchmark's layout boxes (IoU ≥ 0.5). Most of the others are not misplaced: the model gives each line of the numbered conclusions (items 20–29) its own box, while the benchmark puts each pair of lines in one box.
[{"bbox": [940, 14, 977, 26], "category": "header", "text": "8 / 29"}, {"bbox": [409, 75, 590, 92], "category": "title", "text": "การสรุปความจากภาษา"}, {"bbox": [129, 97, 870, 169], "category": "text", "text": "คำสั่ง ในแต่ละข้อจะประกอบด้วยเงื่อนไขหรือข้อมูลที่กำหนดให้ และข้อสรุปเป็นคู่ ๆ จากเงื่อนไขนั้นให้ศึกษาเงื่อนไขที่กำหนดให้ก่อน แล้วจึงอาศัยความรู้ที่ได้จากเงื่อนไขดังกล่าวมาใช้พิจารณาข้อสรุปทั้งสองของแต่ละข้อ แล้วทำตอบลงในกระดาษคำตอบ โดยยึดหลักในการทำตอบดังนี้"}, {"bbox": [130, 202, 272, 219], "category": "title", "text": "หลักในการทำตอบ"}, {"bbox": [129, 224, 474, 245], "category": "text", "text": "ตอบ 1. ถ้าข้อสรุปทั้งสองเป็นจริงตามเงื่อนไข"}, ...]
2. Full-page parsing: a Thai nurse-training handout with references
18 of the 20 predicted boxes match the benchmark's layout boxes (IoU ≥ 0.5).
[{"bbox": [497, 43, 883, 77], "category": "header", "text": "เอกสารประกอบการอบรม การดูแลผู้ป่วยบาดเจ็บขั้นสูง รุ่นที่ 13\nAdvanced trauma emergency care for nurse :ATEN"}, {"bbox": [897, 45, 929, 61], "category": "header", "text": "127"}, {"bbox": [177, 95, 723, 156], "category": "text", "text": "สูตรการคำนวณพลังงานแคลอรี่ (25 Kcal x kg.)+(40 Kcal x % BSA)\nตัวอย่างชายไทย นน. 50 กก. Burn 80% TBSA Total Calories/day\n= (25x50)+(40x80)=4450 Kcal/day"}, {"bbox": [179, 182, 362, 198], "category": "title", "text": "10.การพยาบาลด้านจิตใจ"}, {"bbox": [208, 203, 460, 220], "category": "text", "text": "■ มีการประคับประคองด้านจิตใจ"}, ...]
Prompt:
จากรูปภาพกรุณาดึงข้อมูล เปอร์เซ็นต์ของส่วนประกอบแต่ละอย่าง จากรูปภาพเป็นรูปแบบ JSON เป็นภาษาไทย เฉพาะอันที่มีเปอร์เซ็นต์ โดยมี key ตามนี้มะพร้าวน้ำตาลเกลือสมุทรกลิ่นคาราเมลตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม
(Extract the percentage of each ingredient as JSON in Thai, using the keys coconut, sugar, sea salt and caramel flavour.)
Output, score 1.00 (every other open model we tested, including Typhoon OCR V2 0.8B: 0.00):
{"มะพร้าว": "68%", "น้ำตาล": "29%", "เกลือสมุทร": "2%", "กลิ่นคาราเมล": "1%"}
Prompt:
ตามข้อมูลในภาพกรุณาจับคู่รายการระดับความเข้มข้น PM2.5 (จำนวนจำกัด) เกี่ยวข้องดังต่อไปนี้: 'ดีมาก', 'ดี', 'ปานกลาง', 'เริ่มมีผลกระทบต่อสุขภาพ' และ 'มีผลกระทบต่อสุขภาพ' กับค่าที่จะถูกจับคู่ได้แก่ '7 จังหวัด', '0 จังหวัด', '28 จังหวัด', '16 จังหวัด' และ '25 จังหวัด' เลือกรายละเอียดจากเนื้อหาในเอกสารแล้วให้คำตอบกลับมาในรูปแบบ JSON โดยตรง:ตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม
(Match each PM2.5 level, from "very good" to "affecting health", to the number of provinces in the table, and answer in JSON.)
Output, score 1.00 (every other open model we tested, including Typhoon OCR V2 0.8B: 0.00):
{"ดีมาก": "28 จังหวัด", "ดี": "16 จังหวัด", "ปานกลาง": "25 จังหวัด", "เริ่มมีผลกระทบต่อสุขภาพ": "7 จังหวัด", "มีผลกระทบต่อสุขภาพ": "0 จังหวัด"}
5. Handwriting: lecture notes
Prompt: ข้อที่ 4 ในภาพนี้ เป็นข้อความที่เขียนด้วยมือว่าอะไรบ้าง ตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม (What is handwritten in item 4?)
Output, score 0.96 (best of the four other open models we ran on ThaiOCRBench: 0.38; Typhoon OCR V2 0.8B: 0.84):
④ ผูกขาด // ผู้ขายรายเดียว ไม่มีสินค้าอื่นมาทดแทน ex. ไฟฟ้า ประปา
6. Text in a given region: a street sign
Prompt: แบ่งความยาวและความสูงของรูปภาพออกเป็น 1000 ส่วน แล้วช่วยดึงข้อความที่อยู่ในพิกัด [786, 615, 931, 675] ของรูปภาพออกมาให้หน่อย ตอบเฉพาะข้อความตามภาพตรง ๆ อย่างกระชับ โดยไม่ใส่คำอธิบายเพิ่มเติม (Read the text at [786, 615, 931, 675] on a 0–1000 grid; the box is drawn on the image.)
Output, score 1.00 (best of the four other open models we ran on ThaiOCRBench: 0.12; Typhoon OCR V2 0.8B: 0.16):
Usage
The typhoon-ocr Python package (pip install typhoon-ocr), which runs Typhoon OCR 1.5, does not support V2 yet. Use vLLM or Transformers as shown below.
The model was trained on a fixed set of prompts. For full-page parsing, use this prompt exactly as written; it is the prompt from Infinity-Parser2.
PROMPT = """- Extract layout information from the provided PDF image.- For each layout element, output its bbox, category, and the text content within the bbox.- Bbox format: [x1, y1, x2, y2].- Allowed layout categories: ['header', 'title', 'text', 'figure', 'table', 'formula', 'figure_caption', 'table_caption', 'formula_caption', 'figure_footnote', 'table_footnote', 'page_footnote', 'footer'].- Text extraction and formatting: 1) For 'figure', the text field must be an empty string. 2) For 'formula', format text as LaTeX. 3) For 'table', format text as HTML. 4) For all other categories (e.g., text, title), format text as Markdown.- The output text must be exactly the original text from the image, with no translation or rewriting.- Sort all layout elements in human reading order.- Final output must be a single JSON object."""
Although the last line asks for a JSON object, the model returns a JSON list inside a json code block:
[{"bbox": [62, 48, 938, 96], "category": "title", "text": "Annual Report 2025"}, {"bbox": [62, 120, 938, 410], "category": "text", "text": "The company ..."}, {"bbox": [62, 432, 938, 700], "category": "table", "text": "<table>...</table>"}, {"bbox": [480, 950, 520, 968], "category": "footer", "text": "12"}]
Bounding boxes use a 0–1000 grid relative to the page width and height, so they do not depend on the image resolution. These helpers parse the output, convert it to Markdown and map a box back to pixels:
import json, re def parse_elements(raw: str) -> list[dict]: """Parse the model output: a JSON list of elements inside a ```json fence.""" m = re.search(r"```json\n(.*?)(?:\n```|$)", raw, re.DOTALL) text = (m.group(1) if m else raw).strip() try: return json.loads(text) except json.JSONDecodeError: # Output stopped at the token limit: drop the last, incomplete element. cut = text.rfind('{"bbox":') return json.loads(text[:cut].rstrip().rstrip(",") + "]") def to_markdown(elements: list[dict], keep_header_footer: bool = False) -> str: skip = set() if keep_header_footer else {"header", "footer", "page_footnote"} return "\n\n".join(e["text"] for e in elements if e.get("text") and e["category"] not in skip) def to_pixels(bbox, width, height): x1, y1, x2, y2 = bbox return [x1 * width / 1000, y1 * height / 1000, x2 * width / 1000, y2 * height / 1000]
The JSON from the model always contains every element. to_markdown leaves out header, footer and page_footnote elements by default, which is how we evaluated the model. To keep them, pass keep_header_footer=True:
elements = parse_elements(raw) # raw: the text the model returnedmd = to_markdown(elements) # drops header, footer, page_footnote (as in our evaluation)md_all = to_markdown(elements, keep_header_footer=True) # keeps every element
The examples below need PROMPT and the helpers above. They were tested with vLLM 0.33.0 and Transformers 5.15.1. Use the other settings in the code as they are; the benchmark numbers were produced with them.
pip install "vllm>=0.33.0"
On vLLM versions older than 0.33.0, add --enforce-eager to vllm serve, or enforce_eager=True to LLM(...). Without it, vLLM 0.27.1 compiles the model with torch.compile and CUDA graphs, and in that mode we measured about 9 points lower on MDPBench (70.8 against 80.0 for the same weights). We have not tested versions between 0.27.1 and 0.33.0.
vLLM server
vllm serve typhoon-ai/typhoon-ocr2-2b \ --served-model-name typhoon-ocr2 \ --max-model-len 40960 \ --limit-mm-per-prompt '{"image": 1}' \ --mm-processor-kwargs '{"min_pixels": 2048, "max_pixels": 16777216}' \ --default-chat-template-kwargs '{"enable_thinking": false}'# vLLM < 0.33.0: add --enforce-eager
import base64from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY") with open("page.jpg", "rb") as f: image_url = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode() resp = client.chat.completions.create( model="typhoon-ocr2", messages=[{"role": "user", "content": [ {"type": "image_url", "image_url": {"url": image_url}}, {"type": "text", "text": PROMPT}, ]}], temperature=0.0, max_tokens=16384,)elements = parse_elements(resp.choices[0].message.content)print(to_markdown(elements))
vLLM offline
import base64from vllm import LLM, SamplingParams def main(): llm = LLM( model="typhoon-ai/typhoon-ocr2-2b", max_model_len=40960, limit_mm_per_prompt={"image": 1}, mm_processor_kwargs={"min_pixels": 2048, "max_pixels": 16777216}, # enforce_eager=True, # needed on vLLM < 0.33.0 (see above) ) sampling = SamplingParams(temperature=0.0, max_tokens=16384) with open("page.jpg", "rb") as f: image_url = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode() messages = [{"role": "user", "content": [ {"type": "image_url", "image_url": {"url": image_url}}, {"type": "text", "text": PROMPT}, ]}] out = llm.chat(messages, sampling, chat_template_kwargs={"enable_thinking": False}) print(to_markdown(parse_elements(out[0].outputs[0].text))) if __name__ == "__main__": main()
import torchfrom PIL import Imagefrom transformers import AutoModelForImageTextToText, AutoProcessor model_id = "typhoon-ai/typhoon-ocr2-2b"model = AutoModelForImageTextToText.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")processor = AutoProcessor.from_pretrained(model_id) image = Image.open("page.jpg").convert("RGB")messages = [{"role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": PROMPT},]}]inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt", enable_thinking=False, processor_kwargs={"images_kwargs": {"min_pixels": 2048, "max_pixels": 16777216}},).to(model.device) out = model.generate(**inputs, max_new_tokens=16384, do_sample=False)raw = processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)print(to_markdown(parse_elements(raw)))
Questions about a document
To ask a question about a document, send the question as the text prompt in place of PROMPT. The answer comes back as plain text.
Limitations
- The model can fall into repetition loops, repeating text until it reaches the token limit. We saw this most on dense multi-column pages such as Thai newspapers. When that happens,
parse_elements above keeps only the complete elements.
- PaddleOCR-VL-1.6 is well ahead on tables and formulas (OmniDocBench table TEDS 93.38 against 82.94, formula CDM 97.48 against 93.95).
- Key-information extraction is the weakest ThaiOCRBench task at 21.0, and the document-parsing task scores 38.1. Check the output of extraction tasks before using it downstream.
- We did not hold out a validation set for parsing. The SFT checkpoint this model continues from was chosen by MDPBench score, so its MDPBench numbers are likely somewhat optimistic. The GRPO checkpoint is the last step of a fixed schedule and was not selected. The other four benchmarks were not used for any choice.
- Bounding boxes and image size. The model was trained with images of up to 4 MP, and the Usage examples allow up to 16.7 MP. The boxes are about equally accurate at both settings: on 13 digital MDPBench pages larger than 4 MP, the share of predicted boxes matching the benchmark's layout boxes (F1 at IoU ≥ 0.5) was 0.72 at 4 MP and 0.74 at 16.7 MP. Text is more accurate at 16.7 MP: with
max_pixels at 4,194,304 (4 MP), MDPBench drops from 79.8 to 78.2, mostly on Thai pages (81.3 to 78.8) and tables (TEDS 0.634 to 0.558). Keep 16.7 MP unless you need less GPU memory or faster inference. If your boxes look shifted, check your vLLM version first: in vLLM 0.27.1's compiled mode we saw boxes shifted to the right at 16.7 MP. Better box accuracy is a goal for the next version.
- We studied Thai and English only. Chinese is in the training data and in OmniDocBench but was not evaluated on its own, and other languages were not tested.
- Accuracy can depend on the inference engine. With vLLM 0.27.1, the default compiled mode lowered MDPBench by about 9 points in our tests; use vLLM 0.33.0 or newer, or eager mode on older versions (see Usage). Other serving engines (SGLang, llama.cpp, Ollama) have not been tested.
- The model has no safety filtering and can output text that is not in the image. It is meant for reading documents, not for open-ended chat.
Citation
Typhoon OCR V2 technical report: coming soon.
Typhoon OCR 1.5:
@misc{nonesung2026typhoonocropenvisionlanguage, title = {Typhoon OCR: Open Vision-Language Model For Thai Document Extraction}, author = {Surapon Nonesung and Natapong Nitarach and Teetouch Jaknamon and Pittawat Taveekitworachai and Kunat Pipatanakul}, year = {2026}, eprint = {2601.14722}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2601.14722}}
Support