Local library

One pip install, no account, no upload.

Evaluation, development, custom pipelines, and any workload that fits on the machine running it.

Install Docling

What it adds

  • Full document model and every export format
  • Complete offline operation once models are cached
  • Direct access to pipeline internals

What it costs you

  • Throughput is bounded by the local machine
  • No built-in queueing, retries or metrics
  • Model downloads need network access on first run

Docling Serve

The same pipeline behind an HTTP endpoint you operate.

Internal services, team workflows, and applications that should not embed the converter directly.

Deploy Docling Serve

What it adds

  • HTTP API and container images
  • Shared capacity across a team
  • Deployment inside your own network boundary

What it costs you

  • You own scaling, monitoring and upgrades
  • Capacity planning is your responsibility

Managed SaaS

Elastic capacity and managed operations, same document model.

Production applications that need throughput and reliability without running the infrastructure.

Docling for IBM watsonx

What it adds

  • Elastic throughput with queueing and retries
  • Usage controls, quotas and observability
  • Managed runtime updates and regional processing
  • Enterprise identity, support and service commitments

What it costs you

  • Documents are processed by the provider, not on your hardware
  • Request size, rate and concurrency limits apply
  • Retention and processing boundaries are set by the provider's terms

Private / on-prem

Your infrastructure, your models, your update windows.

Regulated environments, data-residency requirements and air-gapped networks.

Docling for IBM watsonx

What it adds

  • Private cloud, data centre or air-gapped deployment
  • Control over models, versions and update windows
  • Logging, secrets and identity integrated with your stack
  • Capacity planning and accelerator guidance

What it costs you

  • Requires infrastructure and operational ownership
  • Accelerator availability shapes achievable throughput