LlamaParse parses your documents in LlamaIndex's cloud. LM-Kit parses them on your hardware, with the most accurate parser you can run yourself.
Both turn PDFs, scans and office files into data an AI can use. The difference is where that work happens, and who sees the documents.
LlamaIndex
LlamaParse parses, extracts, splits, classifies and indexes documents as a hosted service, priced per credit.
LM-Kit
Document Parsing runs inside your application or on your servers, with RAG, extraction and agents in the same engine.
ParseBench is published by LlamaIndex. LlamaParse's two agentic tiers lead overall; LM-Kit is third, first on visual grounding, and first of everything you can host.
The mean of the five ParseBench dimensions: tables, charts, content faithfulness, semantic formatting and visual grounding.
Whether each element comes back in the right place, so an answer can point to its exact source. No method scores higher.
Only the methods you can run on your own hardware. All three LM-Kit levels lead them.
Ranks among all 142 methods on the public board plus LM-Kit's three levels; a dotted row marks methods not shown. LM-Kit scores: the public ParseBench evaluation code and dataset, charts graded without the optional LLM judge. Other methods: parsebench.ai leaderboard, October 3, 2026.
LlamaParse is a strong product, and on some documents a cloud service is the right answer. Here is what it does well.
Its Agentic Plus tier leads ParseBench overall, at 90.2, for 5.6¢ a page.
A hosted service with a 99.9% uptime target: call the API, and the scaling is somebody else's job.
130+ file formats in one API, with Markdown, JSON, XLSX and annotated PDF outputs.
Auto Mode sends each page to the cheapest tier that handles it, which LlamaIndex says saves up to 80%.
Extraction, classification, splitting, indexing and an agent builder sit beside parsing in one platform.
SOC 2, HIPAA and GDPR compliance for its SaaS, and custom BAAs on the Enterprise plan.
LM-Kit brings cloud-class parsing to the documents that cannot leave: on a laptop, a scanning station, or a server with no internet at all.
Parsing runs on machines you administer. There is no processor to disclose and nothing to open in the firewall.
Each element carries its page, box, reading order and confidence: the best visual grounding on ParseBench.
Volume costs hardware you already run, not credits. Parse the archive, then parse it again.
LM-Kit One serves the parser over REST, a command line and MCP; LM-Kit.NET embeds it in-process.
RAG, structured extraction, agents, speech and vision share one runtime, so parsed pages flow straight into them.
One versioned model file, verified on load. Nothing changes under a validated pipeline until you upgrade.
LlamaIndex facts are from its own website and pricing page; scores from the ParseBench leaderboard.
| Topic | LM-Kit | LlamaIndex (LlamaParse) |
|---|---|---|
| Product and deployment | ||
| What you get | A parsing engine you run: SDK and server | A managed document processing service |
| Where it runs | Your machines: edge, servers, air-gapped | LlamaIndex cloud; hybrid or VPC on Enterprise |
| Who receives the documents | No one outside your network | The service, unless deployed in your VPC |
| Works with no internet | Yes, after installation | No, for the hosted service |
| License | Commercial; free to build and evaluate | Commercial service; plans from Free to Enterprise |
| Accuracy on ParseBench | ||
| Best overall score | 85.5 (High), third of all methods | 90.2 (Agentic Plus), first |
| Visual grounding | 85.2, first of all methods | 84.3 (Agentic) |
| Best self-hosted option | 82.3 to 85.5, first of all self-hosted methods | LiteParse, 36.9 (no OCR, no LLM) |
| Output | ||
| Tables | HTML with merged cells | Advanced table extraction |
| Charts as data | Values measured from the drawing | Chart and graph extraction |
| Bounding boxes and reading order | Every element, with a confidence | Layout detection with bounding boxes |
| Formats | Markdown, JSON, HTML, DocLang | Markdown, text, JSON, XLSX, HTML tables, annotated PDF |
| Cost and access | ||
| Pricing model | License; no per-page API charge | Credits per action: 1,000 credits for $1.25 |
| Per page, top tier | Your hardware | 1.3¢ (Agentic), 5.6¢ (Agentic Plus) |
| Interfaces | .NET SDK, REST, CLI, MCP | Web app and API |
| Beyond parsing | RAG, extraction, agents, speech, vision in one runtime | Extract, classify, split, index in the platform |
It usually comes down to one question: may these documents leave your network?
LlamaIndex sells LlamaParse, a commercial document processing service that parses documents in its cloud, or in a customer VPC on its Enterprise plan, billed in credits. LM-Kit is a document parsing engine you run yourself, embedded in .NET with LM-Kit.NET or served to any language by LM-Kit One, with no per-page API charge and no document leaving your machines.
Yes. LlamaParse is a commercial document processing service billed in credits, from a free plan to Enterprise. Its Enterprise plan adds hybrid and VPC deployment. LlamaIndex also offers LiteParse, a separate local parser that extracts text and bounding boxes without an LLM.
On ParseBench, a benchmark published by LlamaIndex, LlamaParse Agentic Plus scores 90.2 overall and LlamaParse Agentic 87.0. LM-Kit scores 85.5 at High, third overall and first of every method you can run yourself, and leads all methods on visual grounding at 85.2.
LlamaParse is a hosted service; its Enterprise plan adds hybrid and VPC deployment. LiteParse, its separate local parser, runs offline but extracts text without OCR or an LLM and scores 36.9 on ParseBench. LM-Kit runs fully offline after installation, including on air-gapped machines.
Yes. LM-Kit One serves the parser over a REST API to any language, from a command line, and to agents through MCP. LM-Kit.NET embeds the same parser in .NET applications.
The full comparison grid, plus the capability pages most relevant to this one. The same engine ships two ways: embedded as LM-Kit.NET, or served as LM-Kit One, the Private AI Application Server.
Document parsing
