Xiaomi: Xiaomi-OCR-0
Xiaomi-OCR-0 is a 0.8B document OCR model from Xiaomi, built on Qwen3.5-0.8B-Base, for text, tables and formulas in documents (96.83 overall on OmniDocBench v1.6, self-reported). Send one page image per request to /v1/chat/completions with prompts such as 'Extract the text in the image.', 'Parse the table in the image into OTSL.' or 'Identify the formula in the image and represent it using LATEX format.'. Served from the upstream weights on Phala GPU TEE.
Supported API shape
input
text · image
output
text
max output tokens
8,192
tools
Not listed
json mode
Supported
Verification
Available evidence. Use a verified client to authenticate the quote, connection, and receipt.
Verify private inferencereceipt
signed receipt + hashes
attestation
gateway report
session
attested upstream
serving provider
read from receipt session
Provider
Phala
GPU TEE
input
$0.03/M
output
$0.03/M
cache read
$0.01/M
context
8K
Performance comparison
Last 72h · UTCMore models
Other private inference routes.
Z.ai: GLM-OCR
GLM-OCR is a 0.9B document OCR model from Z.ai (Zhipu). It reads text, tables and formulas with the prompts 'Text Recognition:', 'Table Recognition:' and 'Formula Recognition:', and extracts fields when prompted with its JSON template format ('请按下列JSON格式输出图中信息:' followed by the JSON keys); response_format json_schema is supported. Page-level parsing uses the glmocr SDK (layout detection on the client). Served from the upstream weights on Phala GPU TEE.
context
8K
input
$0.03/M
$0.01/M cache read
PaddlePaddle: PaddleOCR-VL 1.6
PaddleOCR-VL 1.6 is a 0.9B document-parsing vision-language model from Baidu's PaddlePaddle team (96.33 on OmniDocBench v1.6). It recognizes text, tables, formulas and charts per element with the prompts 'OCR:', 'Table Recognition:', 'Formula Recognition:' and 'Chart Recognition:'. Page-level parsing uses the PaddleOCR doc-parser pipeline (layout detection on the client) pointed at this endpoint. Served from the upstream weights on Phala GPU TEE.
context
8K
input
$0.02/M
$0.01/M cache read
ATH-MaaS: OvisOCR3
OvisOCR3 is a 0.8B end-to-end document OCR model published by ATH-MaaS on Hugging Face, built on Qwen3.5-0.8B. It parses full pages into Markdown with tables and LaTeX formulas, and also handles general OCR and key information extraction (96.74 overall on OmniDocBench v1.6, self-reported). Send one page image per request to /v1/chat/completions with the model's prompts from its model card. Served from the upstream weights on Phala GPU TEE.
context
8K
input
$0.03/M
$0.01/M cache read
Phala: DeepSeek V4.1 Flash 1017 Uncensored
Uncensored variant of DeepSeek V4.1 Flash 1017, built as a lightweight LoRA adapter served unmerged on the official FP8 weights. Refuses far less while keeping the base model's capabilities, tool calling, image input and 1M-token context. Served with SGLang in a TDX-attested enclave with NVIDIA confidential computing. Reasoning is on by default and can be turned off or capped per request.
context
1.0M
input
$0.30/M
$0.006/M cache read