Model reference · open weights

TeleOCR

LLMs StarDoc-AI Vision + text 1 build Open weights 16k dl/mo

TeleOCR is an open-weight language model from StarDoc-AI. TeleOCR (BF16) weighs 2.8 GB; the smallest configuration that runs it is RTX 3060 12 GB.

TeleOCR is a 1.4B parameter open-source Vision-Language Model developed by StarDoc-AI for image-text-to-text document parsing. It supports Chinese, English, and Japanese with a context length of 128,000 tokens and is released under the Apache 2.0 license. The model unifies the parsing of digital and camera-captured documents within a single framework.

Summary of the StarDoc-AI/TeleOCR model card, 2026-10-01

Hugging Face ↗Paper ↗

What it is

Released by	StarDoc-AI
Type	Language models
Task	Vision + text
Parameters (lead)	1.4B
Context	128,000 tokens
Runs with	transformers
Released	2026-08-14
Popularity	16k downloads / month
Weights	2.8 GB (TeleOCR (BF16), file size)
Licence	Open weights

What it runs on

Memory and cards for TeleOCR (BF16)

Weights 2.8 GB (file size) · KV cache 115 MB per 1,000 tokens of context, at 16 bits (vLLM's default for this build; an 8-bit cache halves it) · runtime overhead from 1.9 GB on a small card · context up to 128,000 tokens.

Card	Requests at once 8K tokens each	Requests at once 32K tokens each	Longest single request	Counted memory
RTX 3060 12 GB	7	1	58K	11.6 GB
RTX 4060 Ti 16 GB	11	2	90K	15.4 GB
RTX 3090 24 GB	19	4	all 125K	23.4 GB
RTX 4090 24 GB	19	4	all 125K	23.4 GB
RTX 5090 32 GB	27	6	all 125K	31.0 GB
L40S 48 GB	41	10	all 125K	44.0 GB
A100 80 GB	78	19	all 125K	78.2 GB
H100 80 GB	73	18	all 125K	78.1 GB
RTX PRO 6000 Blackwell 96 GB	90	22	all 125K	93.8 GB
DGX Spark (GB10) 128 GB unified	104	26	all 125K	107 GB
H200 141 GB	137	34	all 125K	138 GB
B200 180 GB	178	44	all 125K	176 GB
2× RTX 3060 12 GB tensor parallel	17	4	all 125K	11.6 GB a card

Memory needed at each load

Requests at once	8K tokens each	32K tokens each
1	5.7 GB	8.5 GB
5	9.4 GB	23.5 GB
8	12.3 GB	34.8 GB
16	19.8 GB	64.9 GB
32	34.8 GB	125 GB
64	64.9 GB	245 GB

On one card, with vLLM's small-card settings (2,048 tokens a step). Cards of 70 GB and more reserve more per request and more overhead — each row above uses its own card's settings.

Estimates, not measurements, checked against published vLLM startup logs. The weights are the build's file size; the cache is calculated from its config (grouped-query attention); the overhead is an estimate of vLLM's own memory with that card's default settings. "Requests at once" is how many requests of that length vLLM admits — its reservation at full length, with --max-model-len set to that length; requests that stay shorter fit more. "Longest single request" is the most one request can hold there: below the model's maximum, vLLM starts only with --max-model-len set at or under it. "Counted memory" is vLLM's default 92 % of what CUDA reports for the card (the DGX Spark: about 100 GiB of its shared 128 GB). A tensor-parallel split pools the cards' memory and speeds each token up, at the cost of the link between them; a layer split (llama.cpp) holds more but does not make one request faster. Assumes vLLM 0.10 or later.

From the model card

What StarDoc-AI says about TeleOCR

Read the model card

TeleOCR: Navigating Document Parsing Across Digital and Camera-Captured Documents

🔥 News

2026/09/10 - We have renamed NaviDC-OCR to TeleOCR, and all subsequent model iterations will be developed and released under the TeleOCR version.
2026/09/01 - We noticed that EMNLP 2026 is hosting the Dr.DocBench Challenge, a document parsing competition. We evaluated NaviDC-OCR with its native weights, achieving better results than MinerU 2.5 Pro and PaddleOCR-VL 1.6. Detailed results are shown below dr.docbench-challenge. We welcome the use of NaviDC‑OCR for competitions. Going forward, we will continue to deliver competitive parsing models for the community.
2026/08/29 — Thanks to Nandraj for the GGUF conversion and llama.cpp support! 🔗 NaviDC-OCR-GGUF
2026/08/17 — NaviDC-OCR model weights and technical report have been released.

📖 Introduction

TeleOCR is a lightweight (~1.2B parameters), open-source Vision-Language Model designed specifically for document parsing.

Unlike existing methods that mainly target either digital documents or camera-captured documents, TeleOCR unifies both scenarios within a single framework.

Compared with previous document parsing models, TeleOCR introduces

Multi-node Consensus Voting (MCV) for automatic pseudo-label generation
Geometry-aware document modeling for camera-captured documents
Curvature-Guided Douglas-Peucker Sampling (CGDP)
Image-to-image self-verification for automatic data refinement
Progressive four-stage training pipeline
Content-Structure Decoupled Learning for tables and formulas

These techniques enable TeleOCR to achieve state-of-the-art performance on both digital and camera-captured document benchmarks while remaining lightweight enough for practical deployment.

📊 Experimental Results

TeleOCR achieves state-of-the-art performance on multiple public document parsing benchmarks.

Layout Visualization of Distorted Documents

To evaluate the model's ability to understand complex document deformations, we conduct a visual evaluation on the public dewarping datasets DocUNet and DIR300, with representative results shown in Figure. TeleOCR directly performs layout and content parsing on distorted documents without dewarping preprocessing or a dedicated rectification model, demonstrating robust parsing under complex geometric deformations.

Dr.DocBench Challenge

模型	overall ↑	Text edit ↓	formula cdm ↑	Table teds ↑	order edit ↓
Specialized VLMs
TeleOCR	67.96	0.1903	0.02	64.97	0.398
Mineru 2.5 pro	62.26	0.3402	0.04	67.75	0.356
OvisOCR2	59.25	0.3883	0.00	61.59	0.3791
PaddleOCRvl 1.6	55.11	0.4364	0.21	51.34	0.412

OmniDocBench v1.6

Model Type	Methods	Param	Overall ↑	Text Edit ↓	Formula CDM ↑	Table TEDS ↑	Table TEDS-S ↑	Read Order Edit ↓
Specialized VLMs	TeleOCR	1.2B	96.87	0.027	96.36	97.05	98.52	0.122
	OvisOCR2	0.8B	96.58	0.025	97.53	94.76	97.16	0.111
	PaddleOCR-VL-1.6	0.9B	96.33	0.033	97.49	94.76	97.11	0.127
	MinerU2.5-Pro	1.2B	95.75	0.036	97.45	93.42	95.92	0.120
	GLM-OCR	0.9B	95.22	0.044	97.18	92.83	95.39	0.133
	PaddleOCR-VL-1.5	0.9B	94.87	0.038	96.69	91.67	94.37	0.130
	HunyuanOCR-1.5	1B	94.74	0.033	97.49	94.76	97.11	0.127
	PaddleOCR-VL	0.9B	94.11	0.040	95.70	90.65	93.74	0.135
	Youtu-Parsing	2.5B	93.68	0.044	93.45	92.02	95.00	0.116
	Logics-Parsing-v2	4B	93.27	0.041	95.47	88.42	91.98	0.137
	FireRed-OCR	2B	93.20	0.037	95.27	88.04	91.06	0.131
	MinerU2.5	1.2B	92.98	0.045	95.59	87.88	91.47	0.130
	OpenDoc-0.1B	0.1B	90.64	0.049	92.93	83.88	87.45	0.140
	dots.ocr	3B	90.50	0.048	89.12	87.18	90.58	0.138
	DeepSeek-OCR 2	3B	90.17	0.050	91.59	83.89	87.75	0.144
	HunyuanOCR	1B	89.87	0.089	87.44	91.01	93.23	0.171
	Dolphin-v2	3B	89.34	0.069	90.53	84.40	87.44	0.150
	OCRVerse	4B	88.44	0.063	89.14	82.44	86.27	0.163
	MonkeyOCR-pro-3B	3B	88.43	0.074	88.33	84.35	88.62	0.189
General VLMs	Ovis2.6-30B-A3B	30B	93.62	0.035	94.93	89.44	92.40	0.135
	Gemini 3 Pro	--	92.85	0.064	95.83	89.15	92.96	0.165
	Gemini 3 Flash	--	92.58	0.066	95.03	89.29	93.51	0.173
	Qwen3-VL-235B	235B	89.78	0.063	92.53	83.07	86.75	0.166
	GPT-5.2	--	86.52	0.114	88.00	82.95	87.93	0.193
	InternVL3.5-241B	241B	83.61	0.130	89.52	74.35	79.78	0.215

Wild_OmniDocBench

Model Type	Methods	Param	Overall ↑	Text Edit ↓	Formula CDM ↑	Table TEDS ↑	Table TEDS-S ↑	Read Order Edit ↓
Decoupled VLMs	TeleOCR	1.2B	88.53	0.1173	88.26	89.05	92.14	0.2011
	PaddleOCR-VL-1.6	0.9B	87.36	0.1369	88.42	85.76	90.14	0.2057
	MinerU2.5-Pro	1.2B	87.33	0.1362	90.15	85.46	90.12	0.2013

Quoted from the model card on Hugging Face — the full card is behind the Hugging Face link above.