HunyuanOCR-1.5
Release date: July 7, 2026, according to Tencent’s official announcement. (github.com)
Positioning and capabilities
HunyuanOCR-1.5 is a lightweight, end-to-end vision-language model specialized in optical character recognition (OCR). It unifies document parsing, text detection and recognition, information extraction, text-in-image translation, and multi-image document understanding. The published model card lists approximately 1 billion parameters. (github.com)
The release extends work on low-resource-language and ancient-script OCR, chart and table parsing, and text-centric questions across multiple images. Tencent reports that its training recipe extended the context window to 128K and maximum image resolution to 4K; these training specifications should not be assumed to be guaranteed limits for every deployment. (github.com)
Uses and deployment
Potential uses include extracting and structuring text from documents, tables and formulas; translating text in images; and answering questions about document images. Tencent provides model weights, inference and training resources. Deployment options include Transformers and vLLM, and PC deployment through llama.cpp with converted GGUF weights. DFlash speculative decoding is an optional acceleration approach; actual speed depends on the deployment and workload. (github.com)
Limitations and license
OCR output can contain recognition errors or hallucinated text, so important results should be checked against the source image. Tencent’s paper notes that faithful generation and hallucination suppression remain areas for further research. (arxiv.org)
The model is distributed under the Tencent Hunyuan Community License Agreement, not an unrestricted license. Its terms require certain users with more than 100 million monthly active users to obtain a separate license; review the agreement before deployment or commercial use. (github.com)
Pricing
As checked on November 16, 2026, model-specific official API input and output prices could not be verified on either Tencent’s mainland China or international sites. No CNY or USD amount is asserted. The published weights can be self-deployed, but infrastructure and operation may incur costs. Tencent’s general TokenHub billing documentation does not establish a price for this specific model. (cloud.tencent.com)