Upstage

Upstage is an AI company that offers the Solar family of language models and document-processing products, Document Parse and Information Extract, reachable through an API.

Document Parse converts PDFs, scanned images and office documents into HTML or Markdown, preserving tables, charts and section hierarchy, according to the official site. A mode called Enhanced reads harder visual elements, such as complex tables, checkboxes, charts and diagrams.

Information Extract returns user-defined fields as JSON, with location coordinates and confidence scores, from a schema and without prior training.

The Solar family brings together language models. The console documentation lists models in the Pro and Mini lines with a 512K-token context and up to 128K output tokens, in English, Korean and Japanese. The console also documents text embeddings generation.

The three products complement each other: Document Parse prepares the content, Information Extract structures the fields and Solar processes the resulting text.

Upstage is for companies and development teams that need to turn documents into structured data and feed AI applications with the text of those documents.

The most direct use case is processing documents with complex layouts, such as contracts, forms and reports with tables and charts. The list of formats Document Parse accepts includes JPEG, PNG, BMP, TIFF, HEIC, PDF, DOCX, PPTX, XLSX, HWP and HWPX, which serves teams that receive files from the Korean word processor Hangul.

The limits declared in the documentation guide planning. The synchronous API accepts up to 100 pages per file and the asynchronous one up to 2,000 pages. The languages currently supported are Korean, English and Japanese, and hanzi and kanji are in beta.

RAG and agent teams can combine Document Parse with the Solar models. Information Extract has an integration with LangChain.

Before starting, the team needs an API key from the console and, for Information Extract, the schema of the fields it wants to extract.

The main features of Upstage are Document Parse, Information Extract and the Solar family of language models.

• Document Parse: converts PDFs, images and office documents into HTML or Markdown, preserving tables, charts and section hierarchy. It accepts JPEG, PNG, BMP, TIFF, HEIC, PDF, DOCX, PPTX, XLSX, HWP and HWPX.
• Enhanced mode: reads complex tables, checkboxes, charts and diagrams, according to the company blog.
• Processing limits: 100 pages per file on the synchronous API and 2,000 on the asynchronous one, with files up to 100 MB and 500 MB respectively, according to the documentation.
• Information Extract: schema-based extraction, with JSON output, location coordinates and confidence scores, without training.
• Solar: language models, with the Pro line described by the company as a model for agents and the Mini line aimed at high-volume agent workloads. The console list shows a 512K-token context and output of up to 128K.
• Embeddings: the console documents text embeddings generation.

The languages of Document Parse and of the current models are Korean, English and Japanese.

Upstage is worth it for teams that process many documents with tables and varied formats in Korean, English or Japanese, and less so for teams that need other languages.

Pros:

• Document Parse accepts image and office formats, including HWP and HWPX, which covers files common in South Korea.
• The HTML or Markdown output preserves tables and section hierarchy, which helps teams feeding RAG applications.
• Information Extract returns JSON with coordinates and confidence scores and works from a schema, without training.
• The three products, parsing, extraction and Solar models, sit in the same console and the same API.

Cons:

• The languages currently supported are Korean, English and Japanese, and hanzi and kanji are still in beta.
• The synchronous API is limited to 100 pages per file, and larger files need the asynchronous API, which accepts up to 2,000.
• Models move through lifecycles: according to the documentation, a legacy Solar Mini alias will be discontinued on October 30, 2026, and migration is up to the team.
• The performance figures Upstage publishes come from the company's own internal benchmarks.

The alternatives to Upstage depend on the product, because it combines document processing and language models.

• Amazon Textract: an AWS service that extracts text, tables and forms from scanned documents.
• Azure AI Document Intelligence: a Microsoft service for extracting text, tables and field-value pairs from documents.
• Google Document AI: a Google Cloud service with document processors for data extraction.
• LlamaParse: a document-reading service from LlamaIndex, aimed at feeding RAG applications.
• OpenAI: offers language models through an API that accept images as input, a general-purpose approach rather than a dedicated document-parsing service.

The central difference from Upstage is the package. The cloud services are tied to their provider's ecosystem, and Upstage brings parsing, schema-based extraction and its own models together in the same console. For processing HWP and HWPX files, confirm each alternative's support in its documentation before choosing.